In the conventional distribution network dispatching mode, the forecasting error and real-time fluctuation are regulated by controllable distributed power supply, energy storage and flexible load, and the load values are predicted roughly, which makes the load forecasting often inaccurate. Moreover, using controllable distributed power supply, flexible load or energy storage device to balance load fluctuation of distribution network will result in large fluctuation cost and reserve cost. A multi-objective optimal dispatching method for active distribution network based on cluster load prediction is proposed in this paper. The fuzzy clustering method is adopted to cluster the load of the active distribution network, and the extreme learning machine is utilized to predict the load cluster. Based on the predicted value, the day-ahead dispatching is firstly carried out with the lowest active power dispatching cost, and then modified on the basis of day-ahead dispatching. Real-time dispatching is carried out with the objective of the minimum controllable distributed output correction, the minimum energy storage output correction, and the minimum flexible load correction.