임베디드 기반 스마트온실 인공지능 환경 제어 신뢰성 및 범용성 연구

Abstract

Smart greenhouses use embedded systems to regulate crop environments. This study tests whether an artificial intelligence control model can operate reliably and in a generalizable way across heterogeneous facilities in Korea. The same LSTM based environmental control model is applied to two greenhouses in the same region: one large scale facility and one small to medium scale facility. Training data come from leading farms collected through the Rural Development Administration four year project from 2018 to 2021 entitled Development of Agricultural Big Data Acquisition and Productivity Enhancement Models. The inputs are restricted to nine signals: outside air temperature, wind direction, wind speed, solar radiation, cumulative solar radiation, precipitation, ventilation temperature setpoint, heating temperature setpoint, and indoor air temperature. Experiments are conducted in May 2025, a month that features frequent changes in heating and ventilation setpoints. Performance is expressed as the difference between the target temperature and the actual indoor air temperature. In the small to medium scale greenhouse, the metrics are MSE 0.032, RMSE 0.179, MAE 0.132, MAPE 0.574, and R² 0.998. In the large scale greenhouse, the metrics are MSE 0.022, RMSE 0.151, MAE 0.113, MAPE 0.484, and R² 0.999. The model operates as a continuous closed loop, rapidly adjusting heating, ventilation, and the target temperature when internal states change. Both greenhouses maintain an average temperature band within ±0.1 °C, supporting appropriate crop growth.