Dynamic UAV Deployment in Multi-UAV Wireless Networks: A Multi-Modal Feature-Based Deep Reinforcement Learning Approach

The use of Unmanned Aerial Vehicles (UAVs) as aerial base stations has attracted increasing research interest in recent years. A key challenge in this field is determining how to deploy multiple UAVs in dynamic environments, particularly where mobile user demands fluctuate. To address this challenge, this paper presents an adaptive UAV deployment scheme in a dynamic multi-UAV wireless network, considering the mobility of UAVs and users, state variability, and adjustable UAV transmission power. By jointly optimizing the UAVs’ operational modes, transmission power levels, and movement strategies, our objective is to achieve a trade-off between minimizing power consumption and maximizing ground user coverage. A Deep Reinforcement Learning (DRL) approach is proposed to address these challenges. To capture the dynamic variations of users and UAVs in the environment, a multi-modal feature state space is designed, consisting of both a multi-channel image and vectors. The image component integrates real-time data on user distribution and the UAV coverage area, while the vectors represent UAV operational modes, position data, and system temporal information. These multi-modal features are processed using a combination of Convolutional Neural Networks (CNNs) and Multilayer Perceptrons (MLPs) for advanced feature extraction. To enhance training stability and efficiency, the proposed approach updates parameters using the Proximal Policy Optimization (PPO) method. Simulation results demonstrate the effectiveness of the proposed scheme in balancing power consumption and coverage while effectively managing system dynamics.