<p>The rise in greenhouse gas (GHG) emissions due to growing electricity demand has resulted in increasing global warming-related extreme events, contributing to frequent power grid outages. Microgrids are of significant interest among researchers and practitioners for integrating renewable energy and ensuring power supply. Most studies about microgrid planning either focus on microgrid’s grid-connected mode or its islanded mode, which cannot fully reflect the complexity of dynamic transition between two modes during the actual operation of microgrids. In this paper, we conduct long-term expansion planning of interconnected multi-energy microgrids (IMMGs) under both grid-connected and islanded modes. Multiple real-life requirements, such as power resilience and environment protection constraints, are presented to achieve the objectives of minimizing total cost for microgrid expansion planning (MEP), power outage loss and GHG emissions. Taking three IMMGs as an example, a deep reinforcement learning-based framework is developed to optimize the MEP strategies about the type and capacity of power generation and storage units to invest in each microgrid, as well as the time of these investments. The results show that the proposed framework brings about significant power generation costs and GHG emissions reduction. The effectiveness of the framework on cost-effective MEP with resilience improvement is validated considering power system interdependencies.</p>