Deep neural network based control algorithm for maintaining electrical conductivity and water content of substrate in closed-soilless cultivation

One of the most commonly used soilless techniques is a substrate-based cultivation method which uses a substrate containing nutrient solution continuously fed to the root of crop. However, since the crop roots are strongly influenced by various substrate properties, such as water content, electrical conductivity EC (electrical conductivity) and temperature, the substrate status needs to be accurately monitored prior to optimal nutrient replenishment and irrigation. The objectives of this research were 1) to suggest a feedback-neural network (FBNN) model that predict environmental changes in rockwool substrates such as water content and EC, and 2) to apply a reinforcement learning model for irrigation management with tomato cultivation in greenhouse. The FBNN model includes an environmental prediction and an irrigation control optimization model by using a modified feedback cost function. During the network related operations, the calculated cost of output layer is again feed back to the additional layer at each optimization step. The reinforcement learning was designed to find the best combination of the EC levels and the specific volumes of nutrient solution to be injected. In order to apply these control models, we built an environmental monitoring system based on Raspberry Pi 3 board, and the algorithms were installed on the board to determine the behavior of the actuators to adjust the conditions of the injected nutrient solution. Finally, the performance of each developed models are compared with a commercial nutrient controller. The results show that deep neural network based models are successfully applied to the fertigation system, and the performance to maintain EC and water content of the root zone section was better than the conventional controller.