<![CDATA[This study compared Proximal Policy Optimization (PPO) and Deep Q-Network (DQN) for heating, ventilation, and air-conditioning (HVAC) energy–comfort control using an Internet of Things (IoT)-informed simulation. A dataset comprising 2,027,520 observations of temperature, humidity, voltage, current, electrical power, and occupancy was aggregated into five-minute intervals. After preprocessing, 25,544 intervals from 90 calendar dates were retained; 72 dates were used for training and 18 for chronological testing. Both algorithms shared the same state representation, five discrete HVAC actions, reward formulation, and nominal training budget to enable a controlled comparison. PPO and DQN were trained using five independent random seeds across six scenarios, yielding 60 models. Under baseline conditions, PPO consumed 18.995 ± 0.320 kWh of simulated energy, compared with 19.505 ± 0.116 kWh for DQN. Mean comfort deviations were 0.239 ± 0.072 °C and 0.246 ± 0.069 °C, respectively. Paired Wilcoxon tests with Holm correction indicated a statistically significant PPO–DQN difference only in simulated energy. Reward-function and environmental-parameter sensitivity analyses showed that the relative ranking of comfort and return changed across configurations, whereas PPO consistently used less simulated energy. A fixed rule-based thermostat used the least simulated energy but exhibited greater comfort deviation. The findings were limited to the specified, uncalibrated simulator and did not establish measured HVAC energy savings or universal algorithm superiority.]]>
