Given that the current microgrid incorporates highly connected distributed energy sources, the conventional model control methods do not suffice to support complex and ever-changing operating scenarios. This paper proposes a deep learning-based energy optimization method for microgrid energy management in the new power system scenarios. This article constructs a microgrid cloud edge collaboration architecture, which collects interactive network status data through terminal devices and network edge sides. A microgrid energy management model is constructed based on Bi-LSTM attention in the network cloud. And the model is sunk to provide real-time and efficient comprehensive load and power generation prediction output optimal scheduling decisions at the edge of the network, achieving collaborative control of microgrid light load storage. The simulation based on the actual available microgrid data shows that the proposed Bi-LSTM attention energy management model can achieve rapid analysis and optimize decision-making within 7.3 seconds for complex microgrid operation scenarios.