Hybrid modeling-based reinforcement learning for boiler control in data-scarce buildings using a simulation-informed digital twin

초록

Learning-based models capable of accurately reflecting actual measurement environments are essential to predict the energy consumed by a heating, ventilation, and air conditioning (HVAC) system of a building and to develop optimal conservation strategies. However, in existing buildings, developing high-performance energy prediction models is challenging due to limited data characteristics. These include sparse sensor networks and operational records restricted to fixed setpoints, which lack the variability required for deep learning. This study proposes a novel 'loosely coupled' hybrid modeling method, designated as a Simulation-Informed Neural Network (SINN), that enhances prediction performance by integrating calibrated simulation results with scarce measurement data without requiring the installation of additional sensors. The hybrid model was employed to forecast the gas consumption of boilers for the next day, yielding coefficients of variation of the root mean square error (CV(RMSE)) values of approximately 11 %, which indicates high prediction accuracy. Leveraging this hybrid model to ensure physically consistent responses, reinforcement learning was applied to derive an optimal control scenario aimed at reducing energy consumption. The goal was to manage the operating hours of two boilers evenly while maintaining indoor comfort and minimizing gas usage by ensuring efficient and optimal boiler operation. The analysis demonstrated that the optimal control scenario, based on a two-month heating season dataset, could reduce gas consumption by approximately 23 % while achieving the predefined objectives. These results highlight that the proposed hybrid approach provides a practical and cost-effective pathway toward improving the operation of existing buildings where large-scale sensor retrofits are infeasible.

키워드