Desenvolvimento de um sistema de irrigação inteligente: cultivando alface americana com o auxílio de impressora 3D e aprendizado por reforço

Abstract

This work presents a proposal for a smart irrigation system for water management in vegetable gardens, integrating soil moisture sensors, evapotranspiration models, and artificial intelligence techniques. The system utilizes an ESP32 microcontroller for data collection, integrated with a backend server application responsible for data processing and storage, which communicates with an artificial intelligence model for decision-making. The evapotranspiration forecast is estimated using the Penman-Monteith model and meteorological data. Soil moisture is obtained from sensors calibrated based on agronomic parameters, such as field capacity (FC) and permanent wilting point (PWP). The training of the decision agent is carried out in a simulated environment using a Beam Search heuristic, which allows exploring multiple simulation sequences and selecting those with the highest cumulative reward. During this training, the agent learns to adjust the irrigation volume based on environmental conditions and soil moisture. Following this training, the learned policy is applied deterministically, correcting the irrigation volume proposed by the evapotranspiration forecast to maintain soil moisture close to the target level. The results highlight a clear trade-off between the evaluated training approaches. Agent training guided by the Beam Search local heuristic algorithm showed faster convergence with a stable irrigation policy safe from water risks. On the other hand, training via a Softmax policy, although requiring a higher computational cost and a prolonged training time, shows increasing potential and a higher probability of finding the global optimum.