Adaptive robust greenhouse climate control: Combining deep reinforcement learning and economic optimization

Mohamed Mansour, Kiran Kumar Sathyanarayanan, Philipp Sauerteig, Stefan Streif,
Adaptive robust greenhouse climate control: Combining deep reinforcement learning and economic optimization,
Smart Agricultural Technology,
Volume 12,
2025,
101327,
ISSN 2772-3755,
https://doi.org/10.1016/j.atech.2025.101327.
(https://www.sciencedirect.com/science/article/pii/S2772375525005581)
Abstract: Climate control in semi-closed greenhouses is essential for maximizing crop yield and ensuring energy-efficient operation. Model Predictive Control (MPC) is widely used to optimize control inputs based on system dynamics and operational constraints. However, standard MPC performance can degrade under various system uncertainties. While robust and stochastic MPC address these challenges, they often entail high computational costs. Moreover, adapting them to new greenhouse configurations typically requires model reparameterization or implementing adaptive MPC, which limits scalability. In contrast, model-free approaches like Deep Reinforcement Learning (DRL) offer inherent adaptability by learning control policies directly from data, making them well-suited for handling uncertainties. However, DRL methods can require extensive training data and may suffer from instability during learning. To address these limitations while leveraging the strengths of both approaches, we propose a hierarchical control framework that integrates MPC and DRL. The framework integrates an upper-level controller, which performs economic optimization by considering dynamic energy pricing, with a low-level DRL-based controller that ensures robust real-time reference tracking. The DRL-based controller is trained using a two-stage strategy to ensure robustness and adaptability. We demonstrate the controller's robustness through simulations of a semi-closed greenhouse under three scenarios, including actuator failures and environmental fluctuations. Furthermore, we demonstrate the controller's adaptability when deployed in a similar but structurally altered greenhouse. Results indicate that the DRL-based controller maintains stable greenhouse conditions under various challenges, and can easily adapt to different greenhouse setups. This highlights its potential as a reliable and scalable solution for climate control in modern greenhouse farming.
Keywords: Deep reinforcement learning; Hierarchical control; Greenhouse climate control; Transfer learning; Economic optimization