Optimizing Sensor Placement for Air Temperature Measurements in a Naturally Ventilated Greenhouse Using Machine Learning

Highlights The air temperature at each internal point of the naturally ventilated greenhouse was predicted using machine learning. The simplified ML model was tested to minimize the installation of sensors for learning feature collection in machine learning model development. Using the developed machine learning model, this study proposed an optimal sensor location for predicting the internal air temperature of the greenhouse. ABSTRACT. The recent surge in the number of large-sized greenhouses can be attributed to advancements in automation and mechanization facilitated by the development of information and communication technologies (ICT). Establishing an optimal internal growing environment necessitates the installation and maintenance of numerous sensors for internal environmental monitoring and control of the air conditioning system. Achieving this entails the utilization of machine learning (ML) models to determine the optimal sensor locations, ensuring effective monitoring performance with a minimal number of sensors. In this study, ML models, including artificial neural networks (ANN), support vector regression (SVR), and long short-term memory (LSTM), were developed to predict the internal air temperature at 9 points within a greenhouse. The evaluation of optimal sensor locations was based on the performance of these models. The ML models were constructed utilizing internal environmental and external weather data collected from nine locations in an eight-span greenhouse during the summer season. Among the three ML models considered, the LSTM model demonstrated superior accuracy in predicting greenhouse air temperature and was consequently selected. Subsequently, to ascertain the optimal sensor location for air temperature prediction, four LSTM models (Basic LSTM, Simple LSTM-1–3) were developed with a focus on minimizing the number of training features. The optimal sensor locations were determined to be at #5 and #6, located at the center of the greenhouse. Particularly, the #5 location, as identified in the Simple LSTM-3 model (R 2 = 0.982, RMSE = 0.020, P-RMSE = 0.380), is recommended due to its relatively low requirement for training features and high prediction accuracy. Keywords: Air temperature, Artificial neural network, Greenhouse, Long short-term memory, Machine learning, Monitoring, Optimal sensor location, Support vector regression.