Machine learning models for runway configuration optimization to maximize runway capacity

• Main Author: Lam, Andy Jun Guang
• Other Authors: Sameer Alam
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2025
• Subjects: Other
• Online Access: https://hdl.handle.net/10356/182361
• Institution: Nanyang Technological University

The demand for air traffic has been increasing rapidly over the years. Due to the inability of current infrastructure and systems to manage the demand, there is an imbalance between the capacity of the system and the demand, leading to congestion. One key bottleneck of the system is the capacity of the runway systems at airports. To mitigate the imbalance, several air traffic flow management (ATFM) methods have been employed. One of the areas of ATFM identified to manage the capacity at airports is Runway Configuration Management (RCM). RCM refers to the management of runway configurations at airports to maximise runway capacity. Airports featuring multiple runways have the capability to operate in diverse runway configurations, each with its unique setup. Frequent alterations in these configurations result in a decrease in overall runway capacity. Presently, Air Traffic Controllers (ATCOs) heavily rely on their operational experience and predefined procedures ("playbooks") to plan the utilization of runway configurations. While these conventional methods are practical, they lack the capacity to comprehensively address the intricacies of a dynamic runway system under increasing weather uncertainties. The objective of this thesis is to develop a unified framework to solve the RCM problem. This thesis proposes data driven methods that predict metrics that are crucial for RCM and also ultimately select the optimal runway configuration to be used at an uncertain airport environment. The scope of this study includes the prediction of runway capacity and also the capacity impact of runway configuration changes, which ultimately are used as inputs to the runway configuration optimization model. There are three primary contributions of this thesis. The contributions of this thesis rely on a data-driven approach to the problems. The dataset used in this thesis consist of flight positions, flight schedules, runway configurations and meteorological data fron Philadelphia International Airport (PHL) during the months of April 2019 to January 2020. The first contribution of this thesis is the development of a decision-tree based online machine learning model for runway capacity prediction. Capacity estimates are vital for the efficient management of airport operations, being a pre-requisite for the strategic and tactical measures used to mitigate traffic congestion and optimize air traffic operations at airports. Due to advances in data availability, machine learning has been applied in recent works for the capacity prediction problem. However, these machine learning techniques are static models and cannot be refitted to learn the dynamic changes in the runway system. These methods also either did not take into account the curse of dimensionality or fail to consider certain factors that might affect capacity. The curse of dimensionality refers to issues in the models arising from too many variables, which causes the model to be more complex and prone to overfitting. Other factors not included include weather conditions such as precipitation, and consideration of the weight of aircraft. To mitigate these issues, a capacity prediction model is proposed that utilizes innovative feature engineering methods to approximate a set of variables that can better explain the complexities of the runway system. The capacity prediction model utilizes the Adaptive Random Forest, an online machine learning model which is a type of Random Forests which is built with decision trees that are able to learn incrementally. The data-driven capacity prediction model is trained and tested on data from Philadelphia International Airport (PHL). By the incorporation of innovative feature engineering techniques, the models are able to predict capacities that on average deviate by 3 flights from the actual count. Also, the models achieve a Mean Average Percentage Error of 12.05% for arrivals and 13.16% for departures while learning in an adaptive, online manner. The second contribution of this thesis is the formulation of classification models to predict the transition time and capacity impacts of runway configuration changes. The capacity of a runway system is affected by the runway configuration in use and the transition time to change to a new runway configuration. Better prediction of runway configuration transition times can aid air traffic controllers in selecting the runway configuration that minimises delays. A novel data-driven approach is introduced to model the transition times between directional runway configuration changes, derived by using computed features from the flight positional data. Classification models are also formulated to assign the magnitude of transition times and their impact on runway capacity, utilizing features known in the literature, as well as engineered features including weather coefficients and runway complexity. The transition time model is able to identify the instances where the transition times are ‘High’ approximately 92% of the time. Correctly identifying ‘High’ transition times is important as high transition times lead to greater reduction in runway capacity. This is validated with data from PHL, when the predicted transition time is used as a feature input for the capacity impact model, which correctly identifies periods of unfulfilled demand approximately 89% of the time. The third contribution is the introduction of innovative methodologies for addressing the RCM problem, with the objective of selecting optimal runway configuration to maximize the overall runway system capacity. A new approach is presented, employing Deep Reinforcement Learning (Deep RL) techniques that leverage real-world data obtained from operations at PHL. This approach generates a day-long schedule of optimized runway configurations with a rolling window horizon, until the end of the day, updated every 30 minutes. Additionally, the classification model to predict the capacity impact resulting from runway configuration transitions is incorporated, providing feedback into the optimized runway configurations generation. The Deep RL model demonstrates reduction of number of delayed flights, amounting to approximately 30%, when applied to scenarios not encountered during the model’s training phase. Moreover, the Deep RL model effectively reduces the number of delayed arrivals by 27% and departures by 33% when compared to a baseline configuration. Importantly, it accomplishes this without necessitating an increased frequency of runway configuration changes as compared to the baseline scenario. The baseline scenario is the one where the runway configurations are the ones that are selected by ATCOs in reality as given in the dataset.