A machine learning-based optimal deployment approach for UAV-assisted HetNets

Abstract

The use of Unmanned Aerial Vehicles (UAVs) to assist terrestrial HetNets is a promising approach to improving connectivity and coverage for mobile cellular users. However, for this approach to be successful, efficient deployment strategies are crucial to optimize network performance. Our project proposes a strategy for deploying UAV-assisted HetNets that takes traffic conditions into account, using machine learning for traffic forecasting. This strategy takes into consideration the traffic demand and network topology to dynamically deploy the UAV Base Stations (UAV-BSs). Our goal is to obtain simulation results that demonstrate a significant improvement in network capacity and energy efficiency while reducing interference between the UAV-BSs and terrestrial Macro Base Stations (MBSs) and Pico Base Stations (PBSs). This proposed strategy can serve as a guideline for the deployment of UAVs in HetNets, ensuring the efficient utilization of network resources and improving network performance metrics such as throughput and latency, all while enjoying reduced operational costs. A machine learning approach is employed to capture temporal changes in cellular network traffic and predict future data trends. For traffic forecasting, three machine learning models - LSTM, SARIMA, and exponential smoothing are utilized and compared. The results indicate that the LSTM model outperforms the other models, with an RMSE of 0.2411 and an MAE of 0.1669. In contrast, SARIMA had an RMSE of 0.3064 and an MAE of 0.2353, while exponential smoothing had an RMSE of 1.0899 and an MAE of 0.9345. Finally, the forecasted data is mapped to the corresponding sum rate equivalents obtained from systemlevel simulations of the heterogeneous network. For UAV-BS location optimization, PSO was selected for this project because of its simplicity, population-based methodology, lack of requirement for gradient information, and superior accuracy when compared to other optimization algorithms. The variables of PSO represent potential locations for UAVs in applications that involve optimizing the deployment of UAV-assisted HetNets. The results obtained from the PSO optimization are compared with the random deployment of UAV-BSs. It is observed that PSO is superior to random deployment in terms of sum rate, with a difference of 55.376 bits/s/Hz. This difference corresponds to a rate of 553 Mbps for a 4G LTE network using a 10 MHz bandwidth. Additionally, PSO achieves higher energy efficiency by transmitting an additional 86 kbits of traffic per joule of energy compared to random deployment.