Abstract
Greenhouse microclimate forecasting is critical for monitoring crop growth and resource efficiency in controlled environment agriculture. However, the accuracy and reliability of greenhouse microclimate forecasting face significant challenges due to complex spatiotemporal interactions between environmental parameters and their nonlinear dependencies across multiple temporal and spatial scales. To address these challenges and enable robust method development, this thesis recorded diverse microclimate data from three greenhouses, which provided multi-regional, multi-seasonal, multi-temporal resolution and spatial granularity datasets. Supported by these datasets, this thesis developed three forecasting methods and frameworks that integrate deep learning with multi-scale temporal convolutions, physics-informed constraints, and spatiotemporal modelling techniques to improve accuracy and interpretability of greenhouse microclimate forecasting. These frameworks progressively advance from single-parameter temperature forecasting to multi-task learning (simultaneous forecasting of temperature, humidity and light intensity), from purely data-driven to physics-informed approaches, and from single-point forecasting to three-dimensional spatiotemporal modelling.
First, the Adaptive Time Pattern Network (ATPNet) was designed to address the complex temporal pattern challenges in indoor temperature forecasting by incorporating multi-scale dilated convolutions (MD.Conv) and attention mechanisms to capture temporal dependencies across different forecasting horizons. Compared to baseline methods, the Root Relative Squared Error (RRSE) was reduced by 11.87%, 3.26%, and 10.97% for 6-hour, 12-hour, and 24-hour temperature forecasting horizons, respectively, which demonstrated significant improvements in multi-step forecasting accuracy. Second, the Physics-Informed Gated Recurrent Unit Network (PIGRU) was proposed to enhance forecasting interpretability. This model innovatively integrated greenhouse energy balance principles into recurrent neural architectures, achieving improved average temperature prediction accuracy with RMSE reductions of 5.68%, 2.22%, and 5.56%, and demonstrated successful extrapolation in cross-scenario forecasting. Finally, considering the physical coupling between temperature, humidity, and light intensity, the
Multi-task Spatial-Temporal Mamba (MST-Mamba) framework was proposed to investigate the heterogeneous distribution of greenhouse microclimate through selective state space modules, employed soft parameter sharing and spatiotemporal fusion mechanisms to capture dynamic variations across different spatial and temporal scales, achieved superior forecasting accuracy and spatial prediction precision.
This thesis demonstrates significant potential for optimising microclimate control, enhancing prediction reliability across different greenhouse systems, and improving crop uniformity and management effectiveness.