In this paper, a RL framework for optimal operation and maintenance of power grid under uncertain conditions is established, and a method combining Q learning algorithm with artificial neural network is proposed, which is suitable for large systems with high-dimensional state-behavior space. This paper presents the Bellman solution of the reduced power grid, including predictive state management equipment, renewable generators and degraded components. It is proved that RL can really make use of the information collected from predictive state management equipment, so as to help select the optimal operation and maintenance measures on system components. The strategy proposed in this paper provides an accurate solution comparable to the real optimal solution.