Data-Driven Optimization Method for Recurrent Neural Network Algorithm: Greenhouse Internal Temperature Prediction Model

Abstract

We developed an internal environment prediction model for smart greenhouses using machine learning models. Machine learning models were developed by finding certain rules based on the data obtained from the target system and have the advantage of learning various characteristics that are difficult to define theoretically. However, the model accuracy and precision can change according to the model structure (hyperparameters, algorithms, epoch) and data characteristics. In this study, the analysis was performed according to the collected weather data characteristics. The model performance was low when the amount of training data was obtained over less than three days (4320 ea). The model performance improved with an increase in the amount of training data. Model performance stabilized when the training data volume exceeded seven days (10,080 ea). The optimal amount of data was determined to be between three and seven days, with an average model r2 of 0.8811 and an RMSE of 2.056 for the gated recurrent unit algorithm. This study verified the possibility of developing a predictive model for the internal environment of a greenhouse based on weather data from outside. This study is limited to a specific target greenhouse, and further analysis of data from various greenhouses and climates is necessary to achieve global optimization.