Comparative Evaluation of PLSTM and Machine Learning Models for Greenhouse Microclimate Prediction with Insights from Seasonality Analysis

Abstract

Predicting the microclimate within greenhouses, particularly temperature (Ti), is a crucial aspect of agriculture that contributes to sustainable production and effective management of resources. However, traditional methods for microclimate forecasting often struggle with capturing complex temporal dependencies and adapting to seasonal variations. To address these challenges, this study proposes the Power Long Short-Term Memory (PLSTM) model, an advanced deep learning approach designed to enhance prediction accuracy and stability. We evaluate the performance of PLSTM by comparing it with established machine learning models, including Artificial Neural Networks (ANN), Support Vector Machines (SVM), Bagging Trees (BG), and Boosting Trees (BT), using experimental databases collected during two distinct periods (October and March) and their combination from a greenhouse located in Agadir. The evaluation employs metrics such as standard calibration error (SEC), coefficient of determination ($\mathbf{R}^{\mathbf{2}}$), and normalized root mean square error (nRMSE). Results show that PLSTM achieves superior predictive performance, with an nRMSE of 4.61%, SEC of $0.0535^{\circ} \mathrm{C}$, and $\mathbf{R}^{2}$ of 0.9625 on combined seasonal data, outperforming other models. Notably, PLSTM exhibits low error variability, indicating enhanced stability and robustness. The model’s improved accuracy with increased training data further demonstrates its capability to generalize across seasonal variations. These findings highlight the potential of PLSTM as a reliable and effective alternative for real-time greenhouse climate management, offering significant advantages over traditional machine learning techniques.