Placement of Package Delivery Center for UAVs with Machine Learning Salih Safa Bacanli † , Furkan Cimen † , Enas Elgeldawi ‡ , and Damla Turgut † † Department of Computer Science, University of Central Florida ‡ Department of Computer Science, Faculty of Science, Minia University {bacanli, furkan}@knights.ucf.edu, enas.elgeldawi@mu.edu.eg, turgut@cs.ucf.edu Abstract—Commercially available unmanned aerial vehicles (UAVs) are usually more affordable and feasible for easy deployment compared to military-level UAVs in civilian ap- plications. However, having a bounded range limits the use of commercially available UAVs in package dropping scenarios. In this paper, we have generated a synthetic dataset for the scenario in which drones or UAVs are used to drop packages to two neighborhoods. The charging and package pick-up station is located between two neighborhoods. By leveraging the synthetic dataset, the location of the charging station is predicted by machine learning techniques given the package request frequency, package dropping times of the UAV, and targeted package delay for the neighborhoods. The results showed that deep neural networks and support vector regressor are more successful in deciding the charging station location. I. I NTRODUCTION UAVs have found usage in different application scenarios in our era. For example, Amazon has pilot trial locations where the UAVs are used to drop shipping packages under the project called Amazon Prime Air [1]. UAVs can also be used to drop items to places where traveling to the service locations might be time-consuming due to traffic congestion or lack of developed infrastructure, or time-sensitive items such as emergency medical kits. Commercial UAVs have various flight times; however, bat- tery capacity remains the main limiting factor. The maximum battery life for most commercial UAVs is around 30 minutes. This maximum battery life time might be only sufficient for one package drop with full charge in most cases, considering that the UAV needs to return to its starting location after the package drop. The weight that a small to mid-size UAV can carry is also limited, which is around 0.5 to 1 kg. Considering these limitations, we can assume UAV would be dropping one package at a time with every full charge. In this study, we assume a charging location will be sta- tioned between two neighborhoods, and the distance between these neighborhoods has a maximum value. A UAV reaching these neighborhoods has an average service time with a maximum service time limit. Another parameter considered is the frequency of service requests from these neighbor- hoods. We want to find the charging station location between two neighborhoods when the servicing parameters and an acceptable average delay for one of the neighborhoods are given. We use machine learning techniques including random forest, support vector regressor (SVR), deep learning for this task. As there is not enough public real-world data available, we simulated various parameters. Briefly, the contributions of this study can be summarized as follows: • We created synthetic data for training machine learn- ing algorithms through our simulation with random distribution location between neighborhoods and other parameters. This data can be accessed via GitHub 1 page. • Our algorithm can predict the desired distribution center location when parameters are the distance between two neighborhoods, frequency of service requests from each neighborhood, average and maximum service times inside neighborhood, and desired delay times. We also compared various machine learning techniques accord- ing to their prediction performance. The remainder of the paper is organized as follows. In Section II, we review the existing literature. A detailed description of the methodology is given in Section III. The performance evaluation results are discussed in Section IV and the paper concludes in Section V. II. RELATED WORK More companies are starting to invest in drone delivery systems. There are a few example pilot drone delivery applications [1], [2], [3]. The applications of drone deliveries include ground vehicle usage with scenarios such as traveling long distances with trucks and delivering packages with a drone, which is carried by truck [4], [5], [6], [7], [8], [9]. Huang et al. [7], [8] analyzed the application scenario where public vehicles are used in delivery applications with drones. It is notable to specify that drone-truck delivery routing is an NP-hard problem [10]. Cokyasar [11] explored the battery swapping machine locations for the truck-drone delivery systems using mixed- integer nonlinear programming techniques. Our approach is different from the existing literature as it is not using any vehicle to help the drone delivery system. A mixed linear integer programming approach is used in different studies examining UAV routing problems. Mixed- integer linear programming is mostly aimed at maximizing the coverage area, which may lead UAVs to visit more points of interests [12] or UAVs to visit larger areas [13]. G´ omez-Lagos [14] used an integer programming approach 1 www.github.com/cosai/UAVPackageSyntheticData 978-1-7281-8104-2/21/$31.00 ©2021 IEEE GLOBECOM 2021 - 2021 IEEE Global Communications Conference | 978-1-7281-8104-2/21/$31.00 ©2021 IEEE | DOI: 10.1109/GLOBECOM46510.2021.9685951 Authorized licensed use limited to: University of Central Florida. Downloaded on May 31,2022 at 16:10:09 UTC from IEEE Xplore. Restrictions apply.