Abstract As the increasing penetration of distributed renewable energy sources, the inherent fluctuation of their output power aggravates more control burden for the power system. Based on the direct load control, the regulation potential of distributed flexible loads can be developed to relieve this pressure. This paper proposes a real-time two-stage optimization model for multiple thermostatically controlled load groups to smooth the power fluctuation in distribution network. The upper-level evaluates the load regulation capacity and suppresses the net exchange power fluctuation; the lower-level further minimizes the regulation costs and makes the target demand curves for each load group, which can guide the subsequent load power tracking. Besides, within the load group, a proportion aggregation method is applied to reflect its external power regulation characteristics, while the decomposition program assigns the regulation task to each terminal user. For the complex problem established above, a hierarchical and iterative solving strategy is proposed to obtain the approximate optimal solution. The simulation results show that the proposed approach can effectively decrease the net exchange power fluctuation as well as regulation costs.