Limited time for multi-UAV emergency power inspection requires prioritizing targets based on fault probabilities. In order to improve the inspection efficiency, multi-charging stations can be introduced to reduce the impact of insufficient endurance of UAVs. The problem is formulated as a multi-depot multi-visit team orienteering problem and addressed using a genetic algorithm with a soft actor-critic (SAC) model. The algorithm first incorporates two types of local search operators into the evolution process of the traditional genetic algorithm to optimize the selection of multiple UAVs visiting targets and to reduce the flight path distance of the UAVs. Then, a method for dynamically adjusting the parameters of the genetic algorithm using reinforcement learning is proposed. By using the SAC model based on maximum entropy policy learning, during the iteration of the genetic algorithm, dynamically adjusts crossover, mutation rates, and weight distance ratios in chromosome reinsertion based on past learning and population state. Experiments show the algorithm's effectiveness in small and large-scale tests, with ablation experiments validating the local search operator's effectiveness and the superiority of the parameter tuning method. Finally, the algorithm’s efficacy in various emergency scenarios is validated via simulations.