With the integration of distributed generations and controllable loads, the power grid becomes geographically distributed with a time-varying topology. The operation conditions may change rapidly and frequently; thus, management and control of the smart grid pose great challenges on traditional centralized control strategies. This paper proposes a distributed algorithm for energy management in a power grid, while dynamically minimizing the adjustment costs. The objective function is designed to optimize the overall social welfare considering generation suppliers and load users simultaneously. The proposed algorithm integrates consensus algorithm and optimal control algorithm, which requires only information exchanging among neighboring units and enables the sharing of computational and communication burden among distributed local controllers. It is robust to communication failures and adaptive to topology changes. Simulation results of the IEEE 9-bus, 39-bus systems, and a 200-unit system demonstrate the effectiveness of the proposed algorithm and indicate the promising applications to practical power systems.