Meteorological Risk-Aware Route Planning for Low-Altitude Aerial Networks: A Multi-Agent Reinforcement Learning Approach

Abstract

Unmanned Aerial Vehicle (UAV) swarms operating in low-altitude aerial networks face severe safety threats from complex and dynamic meteorological conditions. However, existing route planning studies often oversimplify these weather impacts into static or binary constraints, lacking refined multi-dimensional risk quantification. Furthermore, traditional planning algorithms often incur high computational costs in multi-UAV route planning, making real-time collaborative replanning challenging in dynamic environments. To address these challenges, this paper proposes a novel Meteorological Risk-Aware Multi-Agent Proximal Policy Optimization (MRA-MAPPO) framework. Specifically, we construct a refined fourdimensional spatio-temporal meteorological risk quantification model that transforms diverse weather elements into a continuous risk index. Based on this, a decoupled Multi-Agent Reinforcement Learning (MARL) method is introduced to achieve multi-UAV meteorological risk-aware route planning. In the first stage, an offline global meteorological guidance field is computed to provide a macroscopic prior, effectively overcoming sparsereward deadlocks in massive 3D spaces. In the second stage, each UAV uses this guidance field coupled with local perceptions to perform fast decentralized online inference. Simulations based on Weather Research and Forecasting (WRF) data show that the proposed method effectively mitigates local deadlocks and avoids non-convex meteorological obstacles in the tested scenarios. Compared with traditional and MARL baselines, MRA-MAPPO achieves a high success rate and near-optimal path quality, while requiring millisecond-level inference time for real-time UAV swarm scheduling.