Uncertainties in renewable energy and demand bring significant challenges for microgrids (MGs) to maintain operational efficiency and safety. To address the issue, this paper proposes a forecast-aware safe reinforcement learning (FASRL) strategy for the optimal scheduling of MG under environmental uncertainties. The proposed FASRL strategy is based on the safe reinforcement learning algorithm, which explicitly enforces operational constraints through a safety layer integrated into the policy network. Besides, to enhance the agent’s awareness of future system states, a forecasting model is introduced based on the long short-term memory (LSTM) method, which can accurately predict time-dependent stochastic variables such as renewable generation and load demand. Experimental evaluations are conducted on a simulated MG system with real-world data. The numerical analysis results demonstrate that the proposed FASRL approach significantly reduces power imbalances and improves safety compliance compared to conventional deep reinforcement learning (DRL) methods.