Sustainable Greenhouse Microclimate Modeling: A Comparative Analysis of Recurrent and Graph Neural Networks

Emiliano Seri, Marcello Petitta, Chryssoula Papaioannou, Nikolaos Katsoulas, Cristina Cornaro · 2025

Read Sustainable Greenhouse Microclimate Modeling: A Comparative Analysis of Recurrent and Graph Neural Networks by Emiliano Seri, Marcello Petitta, Chryssoula Papaioannou, Nikolaos Katsoulas, Cristina Cornaro (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-2502-17371.txt.

Introduction

Emiliano Seri and co-authors compare two neural network types for modeling the climate inside greenhouses, published in 2025. Greenhouses fitted with solar panels can grow food and make power on the same land. To keep crops growing well while making power, growers need good forecasts of heat, moisture, and other indoor conditions.

One model type, recurrent networks, is strong on time patterns but cannot show how variables drive each other. The other, a spatio-temporal graph network, models those links as a directed graph. On a simple data set, the recurrent model won by a wide margin. On a richer, later data set with more drivers, the graph model overtook it by a clear margin. Direction clearly helped once the system grew complex.

Catalog record

CREATOREmiliano Seri, Marcello Petitta, Chryssoula Papaioannou, Nikolaos Katsoulas, Cristina Cornaro
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-17371

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Sustainable Greenhouse Microclimate Modeling: A Comparative Analysis of Recurrent and Graph Neural Networks Emiliano Seri, Marcello Petitta, Chryssoula Papaioannou, Nikolaos Katsoulas, Cristina Cornaro Abstract The integration of photovoltaic (PV) systems into greenhouses not only optimizes land use but also enhances sustainable agricultural practices by enabling dual benefits of food production and renewable energy generation. However, accurate prediction of internal environmental conditions is crucial to ensure optimal crop growth while maximizing energy production. This study introduces a novel application of Spatio-Temporal Graph Neural Networks (STGNNs) to greenhouse microclimate modeling, comparing their performance with traditional Recurrent Neural Networks (RNNs). While RNNs excel at temporal pattern recognition, they cannot explicitly model the directional relationships between environmental variables. Our STGNN approach addresses this limitation by representing these relationships as directed graphs, enabling the model to capture both environmental dependencies and their directionality. We benchmark RNNs against directed STGNNs on two 15-min-resolution datasets from Volos (Greece): a six-variable 2020 installation and a more complex eight-variable greenhouse monitored in autumn 2024. In the simpler 2020 case the RNN attains near-perfect accuracy, outperforming the STGNN. When additional drivers are available in 2024, the STGNN overtakes the RNN ( R 2 = 0.905 R^{2}=0.905 vs 0.740 0.740 ), demonstrating that explicitly modelling directional dependencies becomes critical as interaction complexity grows. These findings indicate when graph-based models are warranted and provide a stepping-stone toward digital twins that jointly optimise crop yield and PV power in agrivoltaic greenhouses. † † highlights: Compare RNNs and directed STGNNs for greenhouse microclimate prediction. RNN excels in the simple 2020 scenario. STGNN beats RNN on a complex 2024 greenhouse with eight environmental drivers. Directional graphs help when added variables amplify interaction paths. Result guides when to pick temporal vs. spatio-temporal models. 1 Introduction Modern agriculture faces increasing pressure to optimize land use while reducing its environmental impact. The integration of photovoltaic (PV) systems into agricultural greenhouses has emerged as a promising solution to these challenges, offering dual benefits of food production and renewable energy generation. However, this integration introduces new complexities in greenhouse management, as PV panels directly influence light distribution, temperature patterns, and overall microclimate dynamics. These Agri-PV systems optimize land use by combining solar energy harvesting with protected crop cultivation, offering potential benefits for both energy and food security. Yet the successful implementation of PV-integrated greenhouses requires accurate prediction and control of the internal microclimate to ensure optimal growing conditions while maximizing energy production. While sophisticated physical models incorporating greenhouse thermodynamics, crop physiology, and PV system performance exist [ 1 ] , their complexity often restricts real-time applications and rapid design optimization. This study, conducted within the framework of the European REGACE project, represents an initial step toward developing a comprehensive digital twin for PV-integrated greenhouses. Rather than relying on complex physical models, we explore empirical approaches that employ advanced machine learning to predict greenhouse microclimates. By comparing traditional Recurrent Neural Networks (RNNs) with directed Spatio-Temporal Graph Neural Networks (STGNNs), we investigate whether explicit modeling of environmental relationships and directional influences can improve prediction accuracy while maintaining computational efficiency. RNNs capture temporal patterns but treat each variable independently, whereas STGNNs embed each environmental variable as a node in a directed graph, allowing the model to learn how changes propagate through explicit causal pathways [ 2 , 3 , 4 ] . We analyse 15-min data from two Mediterranean greenhouses in Volos (Greece): GH2 (2020, six variables, \approx 33 k samples across three seasons) and the more complex GH4 (autumn 2024, eight variables including PAR and CO 2 , 4 k training / 1 k test). Both models use identical hyperparameter budgets and a 96-step (24 h) input window, ensuring that performance differences arise from the modelling paradigm rather than tuning effort. A vanilla RNN achieves near-perfect accuracy on GH2, yet the STGNN overtakes it on GH4 ( R^{2}=0.905 vs 0.740), demonstrating that directional graphs earn their keep when additional sensors amplify interaction pathways. This finding addresses the open question of when graph machinery is necessary for greenhouse prediction, complementing earlier empirical studies that relied solely on temporal models [ 5 , 6 , 7 , 8 , 9 , 10 , 11 , 12 ] . It also advances digital-twin research for PV-integrated houses by showing how physics-informed edge sets can improve data-driven predictions without exhaustive search. Effective micro-climate control remains challenging because internal temperature and humidity respond non-linearly to ambient weather, solar radiation and ventilation events [ 13 , 14 ] , and PV panels introduce further heterogeneity [ 15 , 16 , 17 ] . Hybrid, physics-based simulations capture those processes [ 18 , 19 , 20 , 21 , 22 , 23 ] but are often too slow for on-line use; our results show that a graph-augmented neural network can bridge the gap when variable interactions are rich. The paper is structured as follows: Section 2 describes the data used in this study, including data collection methods, preprocessing steps, and the handling of missing values. Section 3 outlines the methodologies employed, offering a concise introduction to deep learning and detailing the architectures and implementation of the RNN and directed STGNN models. Section 4 presents the analysis and results, comparing the performance of the RNN and directed STGNN models in predicting greenhouse temperature across different seasons. Finally, Sections 5 and 6 discuss the findings, their implications for greenhouse management, and potential directions for future research. 2 Data This study draws on two high-frequency (15-min) data sets collected in Volos (Greece) under the REGACE project. GH2 spans an entire year (January to December 2020; 33 446 records) with six environmental variables, whereas GH4 is an autumn 2024 campaign (9 October to 20 November 2024; 5 000 records) that adds PAR and CO 2 sensors. GH2 (2020) – six-variable benchmark The features considered during the analysis are given below. The labels used during the analysis are shown in parentheses: 1. External Environmental Factors: (a) External Temperature (OUT_temp): The ambient temperature outside the greenhouse. (b) External Relative Humidity (OUT_RH): The relative humidity of the air outside the greenhouse. (c) Radiation Level (OUT_rad): The amount of solar radiation received. (d) Wind Speed (OUT_wind_speed): The speed of the wind outside the greenhouse. 2. Internal Environmental Factors: (a) Greenhouse Temperature (G2_temp): The internal temperature within Greenhouse. (b) Greenhouse Relative Humidity (G2_RH): The internal relative humidity within Greenhouse. The inside-outside temperature trajectories for the autumn period in GH2 is shown in Figure 1 . The winter and summer periods for GH2 are shown in 9 and 11 in Appendix A . During the summer period, the greenhouse employs a cooling system to prevent internal temperatures from rising excessively. Spring was excluded because of the high number of missing values. For modelling each season is split chronologically: the first 80 % of records form the training set, the remaining 20 % the test set. Figure 1 : GH2 temperatures inside and outside the greenhouse for the autumn period The directed edges in our graph structure represent fundamental physical relationships between environmental variables, based on well-established thermodynamic and atmospheric physics principles. Our choice of edge directions is motivated by the following physical mechanisms: 1. External Relative Humidity Dependencies: (a) Temperature \rightarrow RH: Controls air’s water vapor capacity through the

Read the complete text