Deep Reinforcement Learning for Adaptive Temperature Control in Greenhouses: A DQN-Based Framework and Experimental Evaluation

Abstract The integration of artificial intelligence into controlled environment agriculture represents a transformative opportunity for addressing global food security challenges while promoting sustainable resource utilisation. This paper presents a comprehensive Reinforcement Learning (RL) framework for optimizing greenhouse temperature management, leveraging data from the second edition of the Autonomous Greenhouse Challenge. We formulate the climate control problem as a Markov Decision Process (MDP) and implement a Deep Q-Network (DQN) agent that learns to regulate temperature through discrete actions: heating, cooling, or maintaining the current state. Our methodology incorporates a sophisticated reward function that balances multiple objectives, including maintaining optimal growing conditions, minimizing energy consumption, and ensuring temperature stability. The agent is trained and evaluated on a subset of the 2019 challenge dataset, specifically utilizing measurements from the Reference team's compartment to ensure data consistency and reliability. Experimental results demonstrate that the RL agent effectively learns an adaptive control policy, achieving an average comfort zone occupancy of 40.1\% across diverse starting conditions while demonstrating appropriate action selection based on environmental state. The agent exhibits particularly strong performance when initialised within the optimal temperature range, achieving 100\% comfort zone occupancy in multiple evaluation episodes. This research contributes to the growing body of knowledge on AI-driven agricultural automation by providing a reproducible methodology, comprehensive performance analysis, and insights into the practical challenges of implementing RL systems in real-world greenhouse environments. The findings suggest that RL approaches offer a viable path toward autonomous greenhouse management with potential benefits for crop yield optimization and energy efficiency.