An access control strategy of multi-type service demands is proposed to minimize the long-run average cost of the operation in smart grids, which exploits the flexibility of response time to relieve the fluctuation of demand loads. By capturing the characters of stochastic distribution of service demands and users’ behaviors, an analytical model based on continue-time Markov control processes is introduced. Combined with performance potentials estimate on sample path, a simulation based policy iteration algorithm is presented, which can alleviate the curse of dimensions induced by the large system state space and find a near-optimal policy with the rapid convergence rate. Simulation results show the effectiveness of the proposed method.