Modern power grids are facing grand operational challenges due to highly intermittent and uncertain renewable energies as well as new types of loads, etc. In recent years, the rapid development of artificial intelligence (AI) technology has brought up new solutions for optimal control problems with high dimension, high nonlinearity and high dynamics. Based on deep reinforcement learning (DRL), a novel autonomous control platform is presented, which can realize online learning and decision making for power system dispatch and control. The target of the proposed control platform is to transform massive real-time measurements directly into control decisions within sub-second. In order to fully demonstrate the feasibility of the "grid mind", autonomous voltage control and line flow control are taken as two examples to formulate the methodology of DRL-based power system dispatch and control problem. Finally, both deep-Q-network and deep deterministic policy gradient algorithms are applied to demonstrate the strong learning capability of DRL agents and their effectiveness through extensive simulation results.