This paper proposes a deep-learning-based scheduling approach for community microgrids that explicitly accounts for building thermal dynamics and customer comfort preferences. Traditional heating, ventilation, and air-conditioning (HVAC) scheduling models are NP-hard and scale poorly, especially for large systems with many buildings. To address this challenge, we develop a dual-encoder deep learning model that predicts building-level HVAC ON/OFF schedules using temporal load and temperature profiles, along with static building thermal parameters. The proposed model is trained in a supervised manner using solutions generated by an optimization-based HVAC scheduling framework, thereby serving as a computationally efficient surrogate for predicting HVAC schedules within a microgrid. The model is trained on samples generated by the optimization-based HVAC scheduling framework and evaluated using precision, recall, and F1-score. The results indicate strong predictive performance.