Online transfer learning (OTL) for accelerating deep reinforcement learning (DRL) for building energy management

Buildings account for over one-third of global energy consumption and emissions, primarily from heating and cooling operations. Intelligent optimisation through predictive controls, such as deep reinforcement learning (DRL), offers significant potential for energy efficiency. However, DRL faces challenges in generalisation and impractical retraining when applied to different buildings, limiting its scalability. Prior online transfer learning (OTL) approaches relied on simulation or rule-based methods but lacked live learning and real-time optimisation. This study proposes an OTL strategy combining autonomous simulation-based DRL policy pretraining with real-time fine-tuning for rapid adaptation to new buildings. Using the Soft Actor-Critic (SAC) algorithm, it was tested on commercial building energy management simulations. Results showed 18%+ reductions in HVAC energy consumption and 8%+ improvement in thermal comfort compared to rule-based and non-transfer DRL baselines. Empirical validation highlights OTL's potential in overcoming DRL's cold start and training burdens, paving the way for broader deployment in sustainable energy management.