Abstract The variability and intermittency of renewable energy and power load bring great pressure to the dispatch of Micro-Grid, especially intra-day dispatch. In order to reduce the intra-day dispatch pressure, this paper proposes a data-driven two-stage day-ahead dispatch model for islanded Micro-Grid. The first stage dispatch model considers multiple demand responses, and applies phase space reconstruction and machine learning to predict renewable energy output and power load. Multi-objective particle swarm optimization algorithm is used to solve the first stage dispatch model. For the obtained Pareto Front, weight multiple objectives by entropy weight method to get the optimal solution. Since the deviation between the predicted value and the actual value can lead to renewable energy curtailment or load loss, the role of the second stage is to predict the renewable energy curtailment and load loss after the first stage dispatch. Firstly, extreme gradient boosting is used to predict when renewable energy curtailment and load loss occur. Secondly, extreme learning machine is used to predict the amount of renewable energy curtailment and load loss at the corresponding time points. Finally, linear programming and mixed integer linear programming are used to solve the second stage dispatch model. By comparative cases analysis, simulation results show that the enhancement of demand response to system efficiency is at the cost of increasing dispatch cost and reducing reliability. In contrast, the proposed two-stage dispatch method using regulation reserve capacity not only reduces dispatch cost, but also improves system efficiency and reliability.