Forecasting Root-Zone Electrical Conductivity of Nutrient Solutions in Closed-Loop Soilless Cultures via a Recurrent Neural Network Using Environmental and Cultivation Information

Abstract

In existing closed-loop soilless cultures, nutrient solutions are controlled by the electrical conductivity (EC) of the solution. However, the EC of nutrient solutions is affected by both growth environments and crop growth, so it is hard to predict the EC of nutrient solution. The objective of this study was to predict the EC of root-zone nutrient solutions in closed-loop soilless cultures using recurrent neural network (RNN). In a test greenhouse with sweet peppers (Capsicum annuum L.), data were measured every 10 s from October 15 to December 31, 2014. Mean values for every hour were analyzed. Validation accuracy (R²) of a single-layer long short-term memory (LSTM) was 0.92 and root-mean-square error (RMSE) was 0.07, which were the best results among the different RNNs. The trained LSTM predicted the substrate EC accurately at all ranges. Test accuracy (R²) was 0.72 and RMSE was 0.08, which were lower than values for the validation. Deep learning algorithms were more accurate when more data were added for training. The addition of other environmental factors or plant growth data would improve model robustness. A trained LSTM can control the nutrient solutions in closed-loop soilless cultures based on predicted future EC. Therefore, the algorithm can make a planned management of nutrient solutions possible, reducing resource waste.

Keywords: black box modeling; environmental factor; long short-term memory; machine learning; sweet pepper.

FIGURE 1

A diagram of a closed-loop soilless culture system and measured data of nutrient solutions and growth environments. Refer to Table 1 for the measured data (1–20).

FIGURE 2

Daily averages of temperature, relative humidity, and radiation in the greenhouse from 15 October to 31 December. Zeros were excluded when radiation was averaged.

FIGURE 3

A structure of a long short-term memory (LSTM). I, input vectors; O, output vectors; C, cell state; h, tanh for input and output activation function; σ, sigmoidal function for gate activation function; t and t–1, current and previous times, respectively. Refer to Table 1 for the input (I) and output (O).

FIGURE 4

Comparisons of predicted and measured root-zone electrical conductivities (ECs) of nutrient solutions using validation (A) and test (B) datasets. Solid and dashed lines represent 1:1 line and regression line, respectively. The average of each 24-long output was presented.

FIGURE 5

Test accuracies of trained long short-term memory (LSTM) algorithms at different time steps (A) and output lengths (B).

FIGURE 6

Chronological comparisons of predicted root-zone electrical conductivity (EC) via trained long short-term memory (LSTM) and measured data from 0:00 to 23:00 on December 25 (A), from 6:00 on December 26 to 5:00 on December 27 (B), from 12:00 on December 27 to 11:00 on December 28 (C), and from 18:00 on December 29 to 17:00 on December 30 (D). Arrows represent the point of 00:00.

FIGURE 7

Root mean square errors (RMSEs) of electrical conductivity (EC) of nutrient solutions. RMSEs separately calculated based on each prediction for 1 h were compared with total validation and test RMSEs.