Quantized Deep Neural Network Based Optimal Control of Greenhouses on a Microcontroller

Growing crops in Controlled-Environment Agriculture (CEA) farms, such as greenhouses and vertical farms, can help in meeting the demands of urban centers and achieving climate goals. Recently, many advanced control techniques like Model Predictive Control (MPC) and its variants have been developed for energy-efficient operation and minimization of resource utilization. However, real-time implementation of these advanced strategies come along with certain computational hardware requirements, thus, increasing the operating costs. In this work, we propose to learn the MPC policy of a greenhouse control by means of a Deep Neural Network (DNN) in order to be implemented on a low-cost microcontroller. Additionally, we use a feedback law to reduce undesired quantization effects. The efficiency of our approach is exemplified by means of a simulation study for greenhouse control.