Data-Driven Greenhouse Climate Regulation in Lettuce Cultivation Using BiLSTM and GRU Predictive Control
Soumo Emmanuel Arnaud ,∗, Marcello Calisti, Athanasios Polydoros · 2025
Read Data-Driven Greenhouse Climate Regulation in Lettuce Cultivation Using BiLSTM and GRU Predictive Control by Soumo Emmanuel Arnaud ,∗, Marcello Calisti, Athanasios Polydoros (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-21669.txt.
Introduction
Soumo Emmanuel Arnaud, Marcello Calisti, and Athanasios Polydoros present a 2025 paper on climate control for lettuce greenhouses. Heating, cooling, and feeding carbon dioxide to a greenhouse eats energy and money. Process models of greenhouses are often too complex or too rough for reliable control.
The team instead trained two kinds of recurrent networks on data from a validated simulator. The networks forecast temperature, humidity, carbon dioxide, and crop growth. Plugged into a predictive controller, they steer heating, vents, and gas injection to save energy while respecting the crop's limits. Both controllers beat a standard predictive baseline. Humidity violations fell from about fifty-five percent of the time to under eighteen percent, and day-night temperature swings stayed small.
Catalog record
| CREATOR | Soumo Emmanuel Arnaud ,∗, Marcello Calisti, Athanasios Polydoros |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 42 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-21669
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Data-Driven Greenhouse Climate Regulation in Lettuce Cultivation Using BiLSTM and GRU Predictive Control Soumo Emmanuel Arnaud ,∗, Marcello Calisti, Athanasios Polydoros Abstract Efficient greenhouse management is essential for sustainable food production in response to a growing global population. However, maintaining optimal indoor climates requires significant energy and resources, making advanced control systems critical for economic viability and environmental sustainability. Traditional greenhouse models are often complex and imprecise, limiting the effectiveness of conventional control strategies. To address these challenges, this study investigates data-driven predictive control methods using Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) neural networks. Our experiments showed that GRU-based predictive control reduced temperature and humidity violations by up to 5% and required 40% less computation time than the LSTM approach, all while maintaining equivalent economic performance and crop yield. These findings demonstrate that GRU-based predictive control offers a more efficient and practical solution for real-time greenhouse climate regulation in precision agriculture. 1 INTRODUCTION Global food security faces significant challenges. The number of undernourished people has been steadily increasing since 2014, reaching an estimated 720 to 811 million in 2020 [ 6 ] . This rise occurs as the world population, currently around 7.7 billion, is projected to reach 9.7 billion by 2050, driving a rapid increase in demand for fresh, nutritious food [ 4 ] . Consequently, food production must increase by approximately 70% between 2005 and 2050 to meet future needs [ 5 ] . At the same time, critical resources such as freshwater [ 7 ] , fossil fuels, and arable land are becoming increasingly scarce, while the impacts of climate change continue to intensify. As a result, producing high-quality, fresh food in a resource-efficient manner is more important than ever. Greenhouse cultivation has emerged as a powerful approach to agricultural intensification. Studies indicate that it can achieve substantial yield increases, typically ranging from 10% to 20%, while simultaneously reducing critical resource inputs by 25% to 35% [ 8 , 10 ] . These considerable benefits have contributed to the accelerated global growth of protected horticulture. Nevertheless, the widespread adoption of greenhouse technologies is constrained by inherent challenges, notably their high energy demands and the significant initial capital investments required [ 9 ] . Greenhouses are essential for crop cultivation, as they provide a controlled environment that shields plants from unpredictable outdoor weather. In today’s world where energy costs are rising and labor and resources are becoming increasingly scarce, efficient and automated greenhouse management is more important than ever. This approach helps maximize crop production while minimizing resource use. However, outdoor weather constitutes a significant disturbance, making it challenging to control the greenhouse climate, develop accurate models, and optimize operations for optimal performance [ 3 ] . A greenhouse is a complex system consisting of two primary components: the cultivated plants and the surrounding controlled environment. This internal environment is influenced by a combination of external weather conditions, the functioning of greenhouse equipment, and the physiological activities of the plants themselves. To understand, predict, and optimize the behavior of such systems, researchers commonly rely on mathematical modeling. These models serve a dual purpose: they offer insights into the underlying physical and biological processes and support the development of control strategies to enhance system performance [ 1 , 2 ] . Traditional greenhouse modeling approaches are typically deterministic and based on first-principles, drawing from established scientific understanding of plant physiology, thermodynamics, and fluid dynamics [ 34 , 33 ] . Although these models can provide detailed and mechanistic insights, they are often complex, featuring nonlinear dynamics and a large number of parameters that have to be manually tuned. This complexity reduces their adaptability to different greenhouse setups or operational conditions. Furthermore, first-principles models are generally limited in their ability to handle uncertainty, as estimating the variability in model parameters requires extensive datasets and computationally demanding statistical methods. While these models use data for calibration, they are not data-driven in the machine learning sense—lacking flexibility to adapt to unseen conditions without manual tuning. Building upon these modeling approaches, automated control systems have been increasingly adopted in greenhouse operations to maintain optimal environmental conditions [ 13 ] . Among advanced techniques, Model Predictive Control (MPC) stands out for its superior performance and energy efficiency relative to traditional methods [ 12 ] . The prediction model forms the cornerstone of any MPC strategy; its accuracy directly influences the quality of future state predictions and, consequently, the effectiveness of control actions. MPC typically employs analytical models—often in state-space form—to forecast system dynamics and determine optimal control inputs over a specified prediction horizon [ 11 ] . However, in practical applications, the development of first-principles models is frequently hindered by challenges such as incomplete system knowledge, complex parameter identification, and computational burdens that are incompatible with real-time operation. This has led to the emergence of data-driven black-box models as a viable alternative within the MPC framework. These models, trained exclusively on system input and output data, offer enhanced adaptability and computational efficiency. However, existing machine learning applications in greenhouse climate control, while promising, often struggle with key limitations, including poor generalization to unseen weather, sensitivity to noisy sensor data, and computational demands that impede real-time use. This research directly addresses this gap by integrating and systematically evaluating LSTM and GRU black-box models within the MPC framework for greenhouse control. The study focuses on assessing their predictive accuracy, robustness against environmental disturbances, and practical feasibility for real-time implementation under typical operational constraints and varying prediction horizons. The rest of the paper is organized as follows: Section II presents a detailed review of related work and current trends in data-driven modeling and control for greenhouse and similar systems. Section III introduces the greenhouse system under study, detailing the dataset and key environmental variables used. Section IV explains the architecture, training, and evaluation of the LSTM and GRU models. Section V describes the integration of these models into the MPC framework, outlining the control objectives, constraints, and implementation strategy. Section VI presents the results of simulation experiments, including model performance metrics and a comparative analysis of control effectiveness. Finally, Section VII concludes the paper by summarizing key findings and outlining future research directions. 2 Related work The growing interest in deep learning for time series forecasting has been driven by its powerful ability to model complex nonlinear relationships and automatically extract high-level features from raw data—capabilities that traditional methods often lack [ 18 , 19 ] . Deep learning models are particularly well-suited for handling large-scale, high-dimensional, and multi-modal datasets, enabling them to uncover intricate patterns and significantly enhance prediction accuracy [ 20 ] . These advantages have made deep learning increasingly attractive for greenhouse modelling. For example, [ 21 ] developed a hybrid model that integrates attention mechanisms, convolutional neural networks (CNN), and LSTM networks to accurately predict temperature, humidity, and CO_{2} levels in edible mushroom greenhouses. Similarly, [ 22 ] demonstrated that GRU can achieve accurate minimum temperature predictions even with a limited set of input variables, outperforming other models. In parallel, the rapid advancement of artificial intelligence and data-driven modeling techniques has led to new and effective control strategies. For instance, [ 23 ] , proposed a robust data-driven model predictive