MILP-Based Imitation Learning for HVAC Control

To optimize the operation of a heating, ventilation, and air-conditioning (HVAC) system with advanced techniques such as artificial neural network (ANN), previous studies usually need forecast information in their method. However, inevitably contains errors all time, which degrade performance HVAC operation. Hence, this study, we propose mixed-integer linear programming (MILP)-based imitation learning method to control an without using order reduce energy cost maintain thermal comfort at given level. Our proposed controller is deep (DNN) trained by data labeled MILP solver historical data. After training, our used real-time For comparison, develop two different methods named forecast-based model predictive (MPC) reinforcement (RL) controls The four verified real outdoor temperatures day-ahead prices Detroit city, MI, USA. Numerical results clearly show that MILP-based better than other terms hourly power consumption, daily cost, comfort. Moreover, difference between optimal almost negligible. These are achieved only end day when have full on weather for day.