Abstract: Reasonable task allocation and inspection routes are crucial for robots to replace engineers in performing inspection tasks in dangerous areas of substations. However, most existing studies focused solely on planning fixed shortest paths for inspecting power transformation equipment, neglecting the variability of equipment inspection times and the heterogeneity of inspection levels. To enhance the effectiveness and flexibility of substation inspections, this study establishes a dynamic inspection path planning model by comprehensively considering the variability of inspection times, the heterogeneity of equipment inspection levels, and the number differences equipments to be inspected. To address the NP-hard of the proposed model, this paper proposes a solution based on the reinforcement learning and multi-agent attention mechanism, which first generates inspection paths using an encoder-decoder framework with an attention layer, and then trains it using an unsupervised neural network. Finally, a substation of China Southern Power Grid is used as an experimental site to validate the model. Compared with the genetic algorithm (GA), Hierarchical Variable Neighborhood Search algorithm (HVNS) , and Adaptive Parallel Memetic Multi-Elite Ant System algorithm (APMMEAS) , the proposed algorithm reduces the path distances by 3.31%, 1.24%, and 1.73%, respectively; reduces the planning time by 17.06%, 16.22% and 13.89%, respectively; and reduces the single inspection costs by 21.22%, 6.86%, and 9.14%, respectively.

Key words: multi-agent, power substation, path planning, reinforcement learning, attention mechanism