Due to the randomness of renewable energy generation and the time series coupling characteristics of energy storage systems, it is necessary to properly model uncertain variables and develop optimization algorithms that can efficiently handle multi-objective problems when constructing economic scheduling models for microgrids. In this context, an efficient multi-time scale dispatching method for microgrids based on deep reinforcement learning and heuristic algorithms is proposed, which can take into account uncertain factors and achieve economic and environmental protection operation. The proposed method optimizes the microgrid from two time scales: day-ahead and intra-day. The day-ahead optimization phase utilizes short-term forecast data for initial decision making to minimize the operating cost. For the intra-day dispatching phase, it utilizes the day-ahead optimization scheme as a reference and revises the day-ahead operation scheme if necessary to cope with the real-time fluctuations of renewable energy. The process of intra-day optimization is decoupled into global and local two phases, the global stage is modeled as a non-convex nonlinear optimization problem and solved using heuristic algorithms, while the local stage is modeled as a Markov decision process and solved using deep reinforcement learning methods. Combining deep reinforcement learning with heuristic algorithms improves the training speed and convergence performance of reinforcement learning, avoiding the difficulty of designing reward functions in complex environments. Finally, the case analysis verifies that the proposed scheme achieves optimization of scheduling cost and computing speed, and is suitable for real-time scheduling of microgrids.