Deep reinforcement learning for adaptive control of heater position and heating power in a smart greenhouse

• The proposed system adaptively adjusts the positions and power of multiple heaters. • The heating system improved energy use efficiency and temperature uniformity. • A simulation model for greenhouse heater dynamics and microclimates was developed. • Proximal Policy Optimization and Gated Recurrent Units were integrated for training. Large commercial greenhouses often struggle with spatial heterogeneity of indoor air temperature, which can decrease the crop sales profit by increasing the energy cost of heating, increasing the labor cost for complicated crop management and harming the sales due to non-uniform crop growth and development. The spatial temperature heterogeneity is caused by both static infrastructural factors and dynamic factors such as sun movement and crop states. While solutions have been developed targeting the static factors, effective countermeasures for dynamic factors remain largely unaddressed, constituting the primary objective of our study. We propose an adaptive heating system that dynamically adjusts the positions and the heating power of multiple heaters based on real-time environmental conditions to control the uniformity of temperature distribution and improve heating efficiency. Our deep reinforcement learning (DRL) approach integrates Proximal Policy Optimization (PPO) for optimizing heating strategies and Gated Recurrent Units (GRUs) for predicting temperature variations. According to the simulated experimental results, our adaptive heating system can reduce the temperature standard deviation by up to 25.51% and decrease energy consumption by up to 34.74% compared to the fixed heating system, while maintaining the target greenhouse average temperature. These findings highlight the potential of DRL in adaptive climate control for smart greenhouses, contributing to more energy-efficient and sustainable agricultural practices.