Energy-efficient greenhouse climate control with diffusion reinforcement learning

Greenhouse agriculture is vital for sustainable food production, yet its high energy demand and resource inefficiency pose significant challenges. Traditional climate control methods often rely on heuristic strategies or suboptimal rule-based systems, leading to excessive energy consumption and operational costs. To address this, we propose a diffusion reinforcement learning framework for resource-efficient greenhouse climate control, optimizing temperature, humidity, and CO₂ levels while minimizing energy use. Unlike conventional deep reinforcement learning, or stochastic policy methods, our diffusion-based approach enhances policy robustness by modeling stochastic environmental dynamics, enabling adaptive decision-making under uncertainty. We validate our method using real-world greenhouse data and simulations, demonstrating superior performance over conventional proportional–integral–derivative method. Simulation results show energy savings of 47.31% (±4.14%) in spring, 45.69% (±4.51%) in summer, 55.54% (±2.06%) in autumn, and 42.92% (±2.29%) in winter, compared to baseline methods while maintaining optimal crop growth conditions. This study advances intelligent control in precision agriculture by integrating denoising diffusion probabilistic models with reinforcement learning, offering a data-driven pathway toward energy-efficient and carbon-neutral greenhouse operations. • Diffusion reinforcement learning for greenhouse climate control. • Integrates generative diffusion models with reinforcement learning. • Cuts energy use by 40–58% across seasonal greenhouse operations. • Boosts crop yield by 10.5% and lowers cost by 35.6%. • Enables data-driven, carbon-neutral controlled agriculture.