An Improved Two-Stage Planning Method for Multi-UAV Collaborative Inspection of Spatially Non-Uniform Power Grids

Autonomous inspection of large-scale power transmission grids using multi-UAV systems faces significant challenges, such as the optimal spatial partitioning due to topological heterogeneity, and the individual trajectory planning due to the geometric sensitivity and data scarcity. To address these issues, this paper proposes a two-stage coordinated planning method. In the proposed method, a distribution-aware reweighted discrete balanced spectral clustering (DAR-DBSC) algorithm is designed to dynamically partition the power grid. Through a dynamic feedback optimization mechanism, it improves workload balance while preserving local spatial affinity structure. In addition, an attention-based deep reinforcement learning solver, named spatial symmetric geometric consistency and per-step smooth entropy annealing algorithm (SSGC-PSEA), is proposed for local trajectory generation. In the proposed SSGC-PSEA algorithm, SSGC extracts geometrically invariant features to improve robustness, while PSEA mitigates gradient shocks to ensure training stability. Furthermore, a "Sim-to-Real" curriculum learning strategy is introduced to overcome real-world data scarcity. Extensive experiments on real-world power grid datasets show that the proposed method outperforms the considered baselines in workload balance, optimization accuracy, and cross-scale generalization.