Precise microclimate control across vertical canopy layers is critical for optimizing crop production in horticulture solar greenhouses, yet existing prediction models fail to capture the complex spatiotemporal heterogeneity of temperature and humidity distributions. Here, we develop a long short-term memory-gated recurrent unit (LSTM-GRU) hybrid deep learning framework that integrates environmental parameters with equipment operational status to predict microclimate conditions at multiple canopy heights (0.2–2.0 m) in Chinese solar greenhouses. The model achieved predictive accuracy with mean squared errors of 1.2℃2 for temperature and 5% for humidity, correlation coefficients (R²) of 0.97 and 0.99, respectively, and 89.63% of temperature predictions within ± 1.0℃ and 94.27% of humidity predictions within ± 5%. Compared with the recurrent neural network (RNN), long short-term memory (LSTM), and gated recurrent unit (GRU) models, the hybrid architecture generally showed the best overall predictive performance across canopy heights. For prediction at the 2 m reference height, incorporating equipment operational status (thermal blankets and ventilation) as input variables reduced prediction errors by 23–55% relative to environmental-only inputs. The results further revealed marked vertical microclimate gradients under different operational conditions. This study provides a data-driven framework for multi-layer microclimate prediction in solar greenhouses and may support future real-time monitoring and environmental control application.
Li et al. (Wed,) studied this question.