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

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.

💡 Research Summary

This paper addresses the pressing need for energy‑efficient climate regulation in greenhouse lettuce production by replacing complex first‑principles models with data‑driven recurrent neural networks (RNNs) embedded in a Model Predictive Control (MPC) framework. Training data were generated through high‑fidelity simulations of the validated van Henten greenhouse model, driven by 20 years of real weather records from Schiphol Airport, Netherlands. The resulting dataset comprises 701,200 fifteen‑minute samples, each containing external disturbances (solar radiation, ambient temperature, CO₂ concentration, absolute humidity), control actions (heating power, ventilation rate, CO₂ injection), and internal states (air temperature, humidity, CO₂ level, lettuce dry‑matter growth). After normalisation and removal of missing values, the data were split into ten folds for cross‑validation, ensuring robust assessment of model generalisation.

Two predictive architectures were investigated: a bidirectional Long Short‑Term Memory network (BiLSTM) and a Gated Recurrent Unit (GRU) network. Both models receive as input a concatenated vector of past disturbances, past control inputs, and past outputs over predefined lag windows (l_d, l_u, l_y). The BiLSTM retains long‑term temporal dependencies through separate input, forget, and output gates plus a cell state, while the GRU merges the input and forget gates into a single update gate and eliminates the explicit cell state, thereby reducing the number of trainable parameters by roughly 30 % and accelerating both training and inference. Hyper‑parameters (hidden units, learning rate, batch size) were kept identical for a fair comparison. Both networks were trained using back‑propagation through time and evaluated with root‑mean‑square error (RMSE) and mean absolute error (MAE) on the held‑out folds. Results showed that GRU achieved prediction errors within 5 % of BiLSTM while requiring significantly less computation.

The trained predictors were integrated into an online MPC scheme. At each control interval (15 min), the RNN forecasts the future trajectory of temperature, humidity, CO₂ concentration, and dry‑matter over a prediction horizon of two hours. The MPC then solves a constrained optimisation problem that minimises a weighted sum of energy consumption (heating, ventilation, CO₂ injection) and maximises expected dry‑matter, subject to biological constraints: temperature 18–24 °C, relative humidity 60–80 %, CO₂ 800–1200 ppm, and smoothness limits on control rates. The optimisation is performed using a sequential quadratic programming solver, with the RNN providing the nonlinear dynamics.

Simulation experiments compared three controllers: (i) a conventional MPC using the original first‑principles model, (ii) an MPC with BiLSTM predictions, and (iii) an MPC with GRU predictions. Performance metrics included humidity violation rate, temperature deviation, total energy use, dry‑matter yield, and average computation time per control step. The GRU‑based MPC reduced humidity violations from 54.77 % (baseline) to 15.45 % and temperature deviations were kept below 2 °C throughout day‑night cycles. Energy consumption dropped by approximately 12 % and dry‑matter increased by about 8 % relative to the baseline. Crucially, the GRU controller required roughly 40 % less computation time than the BiLSTM controller, comfortably meeting a real‑time response requirement of under five seconds per control update.

The study’s contributions are threefold: (1) creation of a large, realistic greenhouse simulation dataset grounded in real weather data, (2) systematic validation of RNN predictors through ten‑fold cross‑validation, and (3) demonstration that GRU models can deliver near‑equivalent predictive accuracy to BiLSTM while offering superior computational efficiency, making them well‑suited for real‑time MPC in agricultural settings. Limitations include reliance on simulated rather than field data, which leaves open questions about sensor noise robustness, model drift, and hardware‑specific nonlinearities. Future work should involve long‑term field trials, incorporation of online learning to adapt to changing plant physiology, multi‑resource optimisation (water, nutrients), and formal stability analysis (e.g., Input‑to‑State Stability) to provide theoretical guarantees for the closed‑loop system.