Deep Learning-Driven Throughput Prediction in 5G for UAV-Assisted Emergency Response

Abstract

Mobile and 5G networks offer high data rates, yet they remain vulnerable to congestion during emergency events, potentially degrading service for users and first responders. Accurate throughput prediction is essential for optimizing resource allocation in such scenarios. This paper proposes a deep learning-driven framework for forecasting LTE/5G throughput and dynamically guiding the deployment of Unmanned Aerial Vehicles (UAVs) as temporary base stations to mitigate network congestion. We evaluate two prediction models using real-world 5G metrics from Chicago and Minneapolis (2022–2024): (1) a hybrid CNN-BiLSTM model that captures spatiotemporal dependencies, and (2) a CNN-Image model that transforms sequential metrics into image representations. Results show that both models achieve high prediction accuracy, with CNN-Image outperforming CNN-BiLSTM in mean absolute error and training efficiency. The predicted throughput is then integrated into a Proximal Policy Optimization (PPO) reinforcement learning strategy to guide UAV placement. In simulated emergency scenarios, the PPO-based approach significantly improves average user  hroughput and service coverage compared to random placement and no UAV support. This work demonstrates the effectiveness of
combining deep learning and DRL for enhancing network resilience in disaster-stricken areas.