Greenhouse irrigation control based on reinforcement learning

Abstract

According to the United Nations, the worldwide population will grow to a vast number of 9 billion people by 2050. As the population keeps increasing, meeting the demand for food has become a tough challenge. Therefore, it is necessary to research and develop strategies on agriculture in order to keep up with demand while maintaining sustainability. Precision Irrigation, a sub-branch of precision agriculture, has gained momentum in modern times. This is an area of study about saving water while maintaining and not impacting the growth of the plant. By manipulating the irrigation schedule, one can keep the soil moisture level at the optimum level without stressing the plant.

The objective of this thesis is to explore the implementation and performance of advance closed-loop control systems using artificial intelligence, such as the actor-critic from reinforcement learning, in a controlled environment to optimize the water schedule. The results will be compared against another closed-loop controller, the On-Off control, and an open-loop controller, the Time-Based Control. Water consumption analysis revealed that closed-loop controllers achieved a 40\% reduction in water use, compared to the open-loop controller. Additionally, the actor-critic controller showed a better response at maintaining the soil moisture level closer to the MAD limit compared to the On-Off and Time-based controls.