Reinforcement learning control method for greenhouse vegetable irrigation driven by dynamic clipping and negative...

Area of Science:

• Agricultural Engineering
• Artificial Intelligence
• Environmental Science

Background:

• Greenhouse vegetable production faces challenges with traditional irrigation, leading to water waste and inefficiency.
• Dynamic environmental conditions and complex management factors necessitate advanced irrigation control.

Purpose of the Study:

• To develop and evaluate an advanced irrigation control method for sustainable greenhouse vegetable production.
• To address limitations of conventional irrigation by utilizing enhanced reinforcement learning.

Main Methods:

• Introduction of the Enhanced Negative-incentive Proximal Policy Optimization (ENPPO) algorithm.
• Integration of dynamic clipping functions and negative incentives for continuous action spaces.
• Utilization of real-time sensor data and historical records for optimal irrigation prediction.

Main Results:

• ENPPO demonstrated superior prediction accuracy, convergence efficiency, and water resource utilization compared to PPO and TRPO.
• Minimized instances of both excessive and insufficient irrigation.
• Promoted enhanced vegetable yield and quality while reducing production costs.

Conclusions:

• The ENPPO algorithm offers a versatile technical solution for intelligent irrigation management in greenhouses.
• This approach has substantial potential for advancing sustainable agricultural practices through optimized resource utilization.