<![CDATA[Indoor farming requires coordinated environmental control while limiting electrical energy use. This experimental case study compares two deployed controller configurations on the same physical platform using separate 24-hour records: a fixed Deep Q-Network (DQN) policy and a hysteresis rule-based controller. The DQN used an 11-feature state, an 11-64-64-8 neural network, eight discrete actuator combinations, and operational action constraints. The deployed control intervals were 10 min for DQN and 5 min for rule-based control. Performance was evaluated using target-band compliance, midpoint mean absolute error (MAE), excursion magnitude, measured actuator electrical energy, and actuator-status telemetry. Temperature compliance was 79.86% for DQN and 77.37% for rule-based control, with midpoint MAE values of 1.10 and 1.28 °C. During lamp-on periods, CO2 compliance was 45.31% and 18.23%, with MAE values of 119.31 and 221.69 ppm, respectively. Relative-humidity regulation was limited by the absence of active dehumidification. Controller-dependent electrical energy was 4.508 kWh in the DQN window and 3.788 kWh in the rule-based window. CO2 gas consumption could not be quantified because no flow meter was installed. Because the deployments differed in date, outdoor conditions, update interval, and auxiliary operational constraints, the findings are interpreted as configuration-level field observations rather than isolated causal effects of the decision algorithms.]]>
Copyrights © 2026
