Abstract
Patrol path planning, as the basis of unmanned intelligent patrol, significantly influences the efficiency and quality of power system surveillance. Consequently, this study investigates a multi scene unmanned intelligent patrol technology for power large area, based on an improved reinforcement learning algorithm. The unmanned intelligent patrol model is designed according to the patrol UAVs, wireless charging piles distributed in appropriate locations, and the targets to be patrolled (i.e., multiple scenes within a large power area). On this basis, the shortest patrol path is taken as the objective function for unmanned intelligent patrol path planning, with constraints including flight time limitations,, speed restrictions, and safety distance constraint. The Q-learning algorithm, a subset of reinforcement learning, is used to search the solution space of the objective function, enabling the UAV to select actions that maximize benefits based on this Q-values, thereby determining the next patrol target point. To enhance search efficiency and optimize the reinforcement learning algorithm, prior environmental knowledge is added as heuristic information during the initialization process of the standard Q-learning algorithm, mitigating the randomness of early exploration. A new reward function is developed to implement collision-free traversal checking within the reinforcement learning framework. Experimental results demonstrate that the patrol paths generated by this technology not only ensure the UAV patrol's safety but also meet the requirements of the shortest patrol path, guaranteeing a coverage rate exceeding 98.5% for all designated patrol targets.