Data-Driven Greenhouse Climate Regulation in Lettuce Cultivation Using BiLSTM and GRU Predictive Control

TL;DR

This work addresses the energy-intensive challenge of lettuce greenhouse climate control by introducing a data-driven predictive control framework that embeds LSTM and GRU predictors inside a Model Predictive Control loop. Training data are generated from a validated greenhouse model and validated via $10$-fold cross-validation, demonstrating superior humidity regulation and proximity to optimal economics, with day–night temperature deviations kept under $2^ vert$C and humidity violations reduced from $54.77\%$ (MPC) to the $15$–$24\%$ range for GRU/LSTM. The GRU-based controller offers up to $40\%$ faster computation than LSTM while achieving higher predictive accuracy, making real-time deployment feasible on embedded hardware. Collectively, the results position GRU-based predictive control as a practical, scalable alternative to classical MPC for energy-efficient, high-yield greenhouse automation in commercial protected agriculture.

Abstract

Efficient greenhouse management is essential for sustainable food production, but the high energy demand for climate regulation poses significant economic and environmental challenges. While traditional process-based greenhouse models exist, they are often too complex or imprecise for reliable control. To address this, our study introduces a novel data-driven predictive control framework using Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) neural networks within a Model Predictive Control (MPC) architecture. Training data were generated from a validated dynamic model simulating lettuce cultivation under various environmental conditions. The LSTM and GRU networks were trained to predict future greenhouse states -- including temperature, humidity, CO\textsubscript{2} concentration, and crop dry matter -- with robustness confirmed via $10$-fold cross-validation. These networks were embedded into an online MPC controller to optimize heating, ventilation, and CO\textsubscript{2} injection, aiming to minimize energy consumption and maximize crop yield while respecting biological constraints. Results showed that both the LSTM- and GRU-based controllers significantly outperformed a conventional MPC baseline. For example, humidity violations dropped from 54.77\% (MPC) to 15.45\% (GRU) and 17.71\% (LSTM), while day-night temperature deviations were kept below $2^\circ\text{C}$. The GRU controller further achieved up to 40\% lower computation time than its LSTM counterpart, confirming its real-time feasibility. Overall, the proposed GRU-driven predictive control approach offers a robust and computationally efficient solution for intelligent greenhouse climate automation under practical operational constraints.

Figures (10)

• Figure 1: An overview of the MPC framework combined with LSTM and GRU neural networks is presented. The MPC controller generates control commands based on predictions from the van Henten van1994greenhouse greenhouse model, external disturbances (solar radiation $d_i$, ambient temperature $d_t$, external CO$_2$ concentration $d_c$, and ambient humidity $d_h$), and reference trajectories. The LSTM and GRU networks are then trained on the data generated by this process. These networks learn to predict future greenhouse states—lettuce dry matter ($y_d$), indoor CO$_2$ concentration ($y_c$), indoor temperature ($y_t$), and indoor humidity ($y_h$)—based on historical data, disturbances, and previous control inputs. The modeling capabilities of the LSTM and GRU units to support the optimization process within the MPC loop is investigated.
• Figure 2: A sample of 3840 data points from the training dataset generated using the MPC strategy. The plot includes disturbances (solar radiation in W·m$^{-2}$, outdoor temperature in °C, outdoor CO$_2$ concentration, and outdoor humidity), control actions (CO$_2$ injection in mg·m$^{-2}$·s$^{-1}$, ventilation rate in mm·s$^{-1}$, and heating system output in W·m$^{-2}$), and greenhouse states (lettuce dry matter, indoor CO$_2$ concentration in kg·m$^{-3}$, indoor temperature in °C, and indoor humidity in kg·m$^{-3}$). Time is expressed in days. These data were used to train the LSTM and GRU models.
• Figure 3: 10-fold cross-validation results shown as side-by-side boxplots of RMSE values for GRU and LSTM models across different input window sizes and batch sizes. Each subplot represents a specific batch size ($8$, $16$, or $32$), facilitating comparison of model performance with respect to input window length. The blue markers and red error bars indicate the mean RMSE and the corresponding 95% confidence interval for each configuration.
• Figure 4: Variation of RMSE with training data percentage for GRU and LSTM models during the generalizability test.
• Figure 5: Prediction performance of the GRU model with window size 18 compared to the actual system response. The plot illustrates the model's ability to follow the true dynamics of the greenhouse climate system.