HVAC SYSTEM
I developed a reinforcement-learning-based HVAC control system that learns optimal temperature-setpoint policies by interacting with a simulated building environment. I started by implementing a physics-informed thermal model and a rule-based controller, then I trained DQN, DDQN, SAC, and PPO agents using OpenAI Gym and Stable-Baselines3. I tuned hyperparameters through random search, grid search, and Bayesian optimization to maximize energy savings while ensuring indoor temperatures stayed within comfort bounds. In my experiments, the PPO agent (with grid-search tuning) performed best—achieving about 41.5 kWh of energy savings, keeping temperature deviations under 3.6 °C, and reducing HVAC costs by roughly $4,150 over the testing period. To prevent overfitting, I incorporated dropout, L2 weight decay, and early stopping, and my comparative analysis across algorithms highlights how careful reward design and parameter selection drive both efficiency and comfort in smart-building control.