Strict Recursive Closed-Loop Evaluation of Greenhouse Air Temperature Forecasting with Generated Exogenous Variables...

Area of Science:

• Agricultural Engineering
• Environmental Science
• Machine Learning

Background:

• Accurate greenhouse air temperature forecasting is crucial for environmental monitoring and control.
• Conventional forecasting evaluations often use unrealistic open-loop or semi-closed-loop settings.
• Future observations are frequently unavailable in real-world deployments.

Purpose of the Study:

• To establish a strict recursive closed-loop forecasting protocol for greenhouse air temperature.
• To analyze exogenous error propagation in multi-horizon forecasting.
• To compare the performance of Liquid Neural Networks (LNN) against other models under realistic conditions.

Main Methods:

• Developed a strict recursive closed-loop forecasting protocol and an exogenous-error propagation analysis framework.
• Utilized multi-source environmental sensing data.
• Compared a Liquid Neural Network (LNN) with LSTM, GRU, and XGBoost models for continuous-time forecasting.

Main Results:

• LNN achieved the lowest one-step prediction error (MAE: 0.4771 °C, RMSE: 0.5996 °C).
• LSTM demonstrated superior long-horizon stability under strict recursive closed-loop conditions (6h, 24h, 48h).
• Exogenous-error analysis identified outdoor temperature and solar radiation as key uncertainty sources.

Conclusions:

• Evaluating greenhouse temperature forecasting requires realistic recursive conditions, not just open-loop settings.
• Model stability and the propagation of uncertainty from exogenous variables are critical factors.
• LSTM shows better long-term forecasting performance in closed-loop scenarios compared to LNN, GRU, and XGBoost.