Quantifying the Energy Floor: Direct Measurement and Replay Buffer Bias in SAC-Based HVAC Control on sbsim
Researchers have quantified the minimum achievable energy cost for HVAC control using Soft Actor-Critic (SAC) on a building simulator, finding it to be $35.51 per day. They identified that initializing the replay buffer with pre-filled transitions, rather than starting empty, significantly increased costs by 4.7%. The study also highlighted that the effective planning horizon is much shorter than typically assumed, and that equipment minimum power, not algorithmic design, is the primary constraint on savings. AI
IMPACT Identifies a key constraint in applying RL to building control, suggesting focus on buffer initialization and effective horizon.