Urban Air Mobility (UAM) has seen remarkable growth over the past decade, especially considering the delivery of goods. Notably, the delivery of medical supplies via Unmanned Aerial Vehicles (UAVs) has emerged as a promising application, which is driven by its societal benefits and has been demonstrated successfully in rural environments. However, expanding these operations to urban areas presents a unique set of challenges, such as high traffic density and the need for online path planning of UAVs. Effective path planning and coordination are of utmost importance to allow for safe and efficient delivery operations in urban environments. Currently, path planning of UAVs is relatively unexplored, and for drone delivery models the trajectories of UAVs are often oversimplified. This research aims to fill the research gap by developing and evaluating an online path planning and coordination mechanism in 3D for a cooperative fleet of autonomous UAVs delivering medical supplies. The Rotterdam Area is selected as a use case, but the model can be adapted to other cities. Utilising an agent-based model formulation, this paper formulates a novel approach for multi-agent pathfinding in 3D and real-world environments, where the travelled paths of UAVs in drone delivery models are computed by taking into account the urban environment and kinematics of UAVs. Windowed Cooperative Safe Interval Path Planning (WC-SIPP) is used as a path planning and coordination mechanism, where kinematic constraints of UAVs and safety separation distances between UAVs are incorporated. Experimental studies show the capability of the algorithm to plan UAVs in different multi-layer urban environments. We establish that including more than one vertical layer is beneficial for UAVs to avoid conflicts by allowing them to change altitudes during flight. Additionally, we show that an increase in the number of layers does not lead to an increase in the delivery time of packages. However, an increase in available layers does come with a decrease in computational performance. Besides, the proposed method is able to plan the trajectories of UAVs and ensure safety separation between UAVs within a dynamic environment, where up to 20 unexpected obstacles are introduced into the environment. Furthermore, we demonstrated that for our operation model, assigning prioritisation based on either mission type or package request time did not influence the delivery time of packages.

Predicting aircraft Take-Off Weight (TOW) has been a long-sought task by aviation stakeholders, especially for operational and regulatory bodies involved in flight planning. Unfortunately, TOW being a sensitive parameter to operational trends and cost indices, aircraft operators tend to keep it confidential. In recent years, Machine Learning (ML) algorithms have achieved increased prediction accuracy and capabilities in the field, provided the availability of TOW data. This paper studies the implementation of gradient boosting algorithms as well as Random Forests to better understand which algorithm is best-suited for aircraft TOW prediction (prior to take-off) solely based on Flight PLan (FPL) and Terminal Aerodrome Forecast (TAF) parameters. The study focused on flights at Amsterdam Airport Schiphol (AMS) for training the algorithms, using an 80-20% train-test split. Between Gradient Boosting Decision Trees (GBDTs), LightGBM, XGBoost, and Random Forests, GBDTs achieved the smallest Mean Absolute Percentage Error (MAPE) with 1.71 and 2.17% on the training and testing datasets, respectively. The most influencing feature proved to be the requested cruise speed, followed by great circle distance between airports, and aircraft type. The model was validated on Paris - Charles de Gaulle Airport (CDG) and Brussels South Charleroi Airport (CRL), proving its independence from airport type. However, the distribution of flights in the training dataset, especially that of aircraft and airline types, proved to be an influencing factor for the model's applicability to other airports. Future work includes expanding the training dataset to all flights in the European network, and introducing trajectory-based features such as aircraft speed intent. With a larger training dataset, neural network algorithms could also be explored. Finally, regarding the improvement of trajectory predictions, it was found that better accuracy of TOW predictions does not suffice and that other operational parameters' effect should be investigated, especially speed profiles.

The Dutch healthcare sector wrestles with rising costs, staff shortages, and increased demand due to healthcare centralization. This, coupled with worsening traffic congestion, underscores the need for efficient solutions like drone-based medical transport. This paper addresses the need for effective fleet management strategies tailored to the unique demands of the healthcare environment. Specifically, it seeks to develop an adaptive strategy that accommodates the network's expansion and the inherently stochastic and urgent nature of pickup and delivery orders associated with medical transport. Utilizing an agent-based model formulation, the paper introduces a novel approach combining adapted temporal sequential single-item auctions for allocation and scheduling with reinforcement learning for drone repositioning to optimize fleet management. Our findings highlight the strategy's consistent efficiency across various demand scenarios, maintaining performance within predefined limits. Notably, the repositioning module significantly enhances the fleet utility and the fraction of served orders, albeit at the expense of increased cost per delivery. Conversely, the reallocation module causes minimal performance improvement. Under heightened stochasticity introduced by urgent orders, the strategy maintains stable costs per delivery while fleet utility and order fulfillment rates decline. Additionally, our investigation underscores the increasing benefits of repositioning in more stochastic scenarios. Moreover, exploring hybrid fleets reveals that while short-range high-payload drones can reduce cost per delivery, they compromise overall fleet utility and order fulfillment rates. Furthermore, we identify the under-utilization of payload capacity in scenarios with orders weighing up to 2 kilograms for a drone with a payload of 10 kilograms.

Air traffic delays have a major impact on the aviation industry, affecting airlines, passengers, and the broader ecosystem. With increasing regulatory and sustainability pressures, accurate delay predictions are now critical, as they enable reductions in contingency and discretionary fuel on flights, lowering total fuel usage. This research aims to develop an explainable supervised learning model to improve existing en route delay predictions, focusing on intercontinental flights from North America to Amsterdam Schiphol Airport. While prior studies have explored flight delay prediction, they have not addressed two critical research gaps identified in this research: the inclusion of day-of-operations features, such as passenger information, aircraft weights, and cost index, and the use of transatlantic flight data for
predictions 90 minutes before departure. To address these gaps, two Gradient-Boosted models, CatBoost and LightGBM, were trained using internal airline, airport, and METAR data. Both models outperformed the airline’s current in-use statistical model, with CatBoost achieving an MAE of 3.44 minutes and RMSE of 4.61 minutes and LightGBM achieving an MAE of 3.43 minutes and RMSE of 4.56 minutes. However, the model’s overall predictive accuracy, as
indicated by the R2 scores, remained relatively low, reflecting the inherent challenges of en route delay forecasting. The most significant performance increase over the current model was observed under adverse weather conditions. Despite these improvements, the test run also showed that the performance of the models deteriorates quickly as significant
differences exist in TAFs and actual weather. An explainability framework was developed to aid end users of the model, providing insights into the model’s decision-making process and helping to build user confidence in its predictions. This research advances en route delay prediction by providing more accurate delay forecasts, particularly in critical weather conditions, and proposes practical improvements to support future studies focused on enhancing model adaptability across diverse operational contexts.

This Master Thesis investigates the possible improvements to the Target Time Management concept to optimize the arrival flows for SWISS International Airlines. The aim is to improve operational performance based on the current model used, as well as prove that Target Time Management constitutes a valuable system to improve operations in a dynamic way. To leverage the dynamic nature of slot assignment, an environment model is created and used as training base for two Multi-Agent Reinforcement Learning algorithms. These two algorithms, Soft-Actor Critic (SAC) and Proximal Policy Optimization (PPO), are then tested against the baseline model currently used in operations at SWISS (based on Mixed-Integer Linear Programming). The four domains to measure the algorithms' performance are passenger connecting time, curfew performance, rotation delay and fairness to other airlines. The algorithms were trained in a simulation environment based on statistical representations of the dynamics of the slot allocation system of EUROCONTROL. They were then tested with new data, where they outperformed a MILP implementation in passenger connecting time and rotation delay metrics (curfew and fairness were comparable in magnitude, since the MILP was slightly unfair for SWISS and RL was slightly unfair for other airlines). PPO was then also tested on the real slot assignment environment hosted by EUROCONTROL and once again compared to a MILP approach. Here, it was found that the improvement in critical passenger connecting time was 5.0 minutes for the MILP, and 5.9 minutes for PPO. Rotation delay was improved by 0.9 minutes by the MILP, and by 4.8 minutes by PPO. PPO also made the highest delays higher and the lowest delays lower, which would require EUROCONTROL or SkyGuide representatives to interpret and make conclusions on fairness and safety. Curfew performance was optimal for both methods. In conclusion, it is proven that Reinforcement Learning techniques can aid the dynamicity of decision-making within Target Time Management. It is also proven that Target Time Management with a dynamic decision making approach can improve operational performance compared to a static one.

Punctuality is a key performance indicator for any airline. Hub-and-spoke airlines are particularly committed to on-time arrivals to guarantee passenger connections. Flights that are delayed at departure need to compensate for the lost time whilst airborne. Because fueling takes place well before scheduled departure, predicted departure delays determine the planned fuel amounts for en-route speed optimization. To prevent unnecessary fuel burn, airlines benefit from highly accurate departure delay predictions. This study aims to extend previous work on airline departure delay forecasting to a dynamic and probabilistic domain, whilst incorporating novel day-of-operations airline information to further minimize prediction errors. Random Forest, CatBoost, and Deep Neural Network models were proposed for a case study on KLM departures from Amsterdam Airport Schiphol between 1 January 2020 and 1 August 2023. The Random Forest model was selected for its superior probabilistic performance and high accuracy in predicting delays between 5 and 25 minutes, for which en-route speed optimization has the largest effect. The departure delay probability distribution forecasts are made at six distinct prediction moments: 90, 75, 60, 45, 30, and 15 minutes before scheduled departure time. At the 90-minute prediction horizon, the model reaches a Mean Absolute Error (MAE) of 8.46 minutes and a Root Mean Square Error (RMSE) of 11.91 minutes. Simultaneously, for 76% of flights, the actual delay is within the predicted probability distribution range. Considering the costs and emissions associated with the decision-making following the departure delay prediction model, this study puts strong emphasis on explainability. Flight dispatchers are therefore provided not only the predicted departure delay but also the main factors impacting the prediction, explaining the context of the flight. The versatility of the model was demonstrated in two shadow runs, where delays caused by familiar and unfamiliar factors were successfully predicted.

To facilitate an increase in air traffic volume and to allow for more flexibility in the flight paths of aircraft, an abundance of decentralized conflict resolution (CR) algorithms have been developed. The efficiency of such algorithms often deteriorates when employed in high traffic densities. Several methods have tried to prioritize certain conflicts to alleviate part of the problems introduced at high traffic densities. However, manually establishing rules for prioritizing intruders is a difficult task due to the complex traffic patterns that emerge in multi-actor conflicts. Reinforcement Learning (RL) has demonstrated its ability to synthesize strategies while approximating the system dynamics. This research shows how RL can be employed to improve conflict prioritization in multi-actor conflicts. We employ the Proximal Policy Optimization algorithm with an actor-critic network. The RL model decides on intruder selection based on the local observations of an aircraft. It was trained on a limited number of conflict geometries in which it was able to significantly reduce the number of intrusions. A conflict prioritization strategy was then formulated based on the decisions taken by the RL model during training. We show that the efficacy of a conflict resolution algorithm that adopts a global solution, the solution space diagram (SSD) in this research, can be improved when utilizing this conflict prioritization strategy. Finally, these results were compared to the performance of a pairwise CR method, the Modified Voltage Potential (MVP). Even though MVP resulted in a smaller number of intrusions compared to SSD with conflict prioritization, the prioritization strategy did reduce the gap between the two CR methods.

For future operations of unmanned aviation, even higher traffic densities than previously seen in manned aviation are expected. Previous work has shown that a vertically layered airspace design performs best at improving safety metrics such as the total number of conflicts and Losses of Separation (LoSs). Furthermore, it has been shown that machine learning techniques are capable of selecting heading ranges for the vertically stacked layers in non-uniform traffic scenarios, in order to reduce the number of conflicts and LoSs compared to uniform structures. These works, however, set structures in an ‘empty’ airspace and do not take into account the necessary vertical deviations to get from one structure to the next. In this work reinforcement learning (RL) agents are used to select layer heading ranges, while taking into account the previous airspace structure. During this dynamic structuring, several challenges arise. First, it is not clear how to reduce the number of vertical conflicts when aircraft move into a new airspace structure. Second, specific structures should be selected that reduce the necessary vertical deviations from the old structure, while still minimising the cruising conflicts for the new traffic distribution. The present work is divided into three experiments. Experiment I focused on analysing the number of conflicts and LoSs that aircraft suffer during vertically moving towards their layer in the new structure. Experiment II tested whether a RL agent is capable of setting an aircraft structure in function of the expected future traffic scenario. Experiment III aimed to show the capability of a RL agent to select airspace structures, while taking into account the previous airspace structures, in order to decrease the number of vertical conflicts. The results of the research show that RL methods are capable of defining airspace structures appropriate for a given traffic scenario. For dynamic reconfiguration, it proved challenging to simulate traffic scenarios that cause an agent to select different structures to prevent the occurrence of vertical conflicts. Under the experimental conditions employed, analytical methods of structure selection performed better in terms of safety.

The Last Hope drone will autonomously find a clear path into the sky from the ground and ascend to an altitude of up to two thousand meters. Within 20 minutes it transmits a call for help with exact location information to rescue operators via the Iridium satellite network...

Current estimates show that the presence of unmanned aviation is likely to grow exponentially over the course of the next decades. Even with the more conservative estimates, these expected high traffic densities require a re-evaluation of the airspace structure to ensure safe and efficient operations. One structure that scored high on both the safety and efficiency metrics, as defined by the Metropolis project, is a layered airspace, where aircraft with an intended heading are assigned to a specific altitude layer. However, a problem arises once aircraft start to vertically traverse between these layers, leading to a large number of conflicts and intrusions. One way to potentially reduce the number of intrusions during these operations is by using conventional conflict resolution algorithms. These algorithms however have also been shown to lead to instabilities at higher traffic densities. As recent years have shown tremendous growth in the capabilities of Deep Reinforcement Learning, it is interesting to see how well these methods perform in the field of conflict resolution. This research investigates and compares the performance of multiple Soft Actor Critic models with the Modified Voltage Potential algorithm during vertical manoeuvres in a layered airspace. The final obtained performance of the trained models is comparable to that of the Modified Voltage Potential algorithm and in certain scenarios, the trained models even outperform the MVP algorithm. Overall, the results show that DRL can improve upon the current state of conflict resolution algorithms and provide new insight into the development of safe operations.