Prediction of the ion concentration of root-zone macronutrients using long short-term memory in closed-loop soilless culture of sweet pepper

Closed-loop soilless cultures have been studied for establishment of sustainable agriculture. To prevent the nutritional imbalance, nutrient solutions should be strictly controlled in the closed-loop condition. Recently, deep learning has been used to draw meaningful results from nonlinear data. Long short-term memory (LSTM) as an example of deep learning, is used to analyze time-series data. The objective of this study was to estimate ion concentration of macronutrients in closed-loop soilless culture using LSTM. In the greenhouse with sweet peppers (Capsicum annuum L.), growth and environmental data in a greenhouse were measured every 10 s from January 12 to April 26, 2018, and the average values for every hour were used for input data. In the closed-loop condition, the best accuracy was R²=0.67 with root mean square error=1.48. By using the trained LSTM, the model could estimate the ion concentration of macronutrients in greenhouse environments. Since the LSTM can be easily applied to analyze other accumulative changes, various applications to predict other plant environment or growth can be developed. The interpolated ion concentrations showed similar changes to those observed in normal cultivation. Therefore, the trained LSTM can also be used to detect a sudden malfunction in ion concentrations.