🏡 NARX Modeling of Two Greenhouses with Different Thermal Curtains

Timothy Denen Akpenpuun, Qazeem Opeyemi Ogunlowo, Wook-Ho Na, Prabhat Dutta, Anis Rabiu, Misbaudeen Aderemi Adesanya, Mohammadreza Narimani, Ezatullah Zakir, Hyeon Tae Kim, and Hyun-Woo Lee*

Journal of Agricultural Engineering · Volume LV · Article 1563 · 2024 · Open access CC BY-NC 4.0

Kyungpook National University · University of Ilorin · University of Tehran · Gyeongsang National University

Read the open-access paper

Why the Curtain Position Matters

Protected cultivation decouples crop climate from the outdoor macroclimate, but heating, ventilation, and screens interact as a nonlinear, time-lagged system. This study models two adjacent single-span strawberry greenhouses that differ only in thermal-curtain position: screens in the R greenhouse (RGH) hang in the roof space; screens in the Q greenhouse (QGH) follow a different hang geometry.

Mohammadreza Narimani contributed to methodology with colleagues at Kyungpook National University and partner labs. MATLAB NARX (nonlinear autoregressive with exogenous inputs) networks, trained with Levenberg–Marquardt backpropagation on normalized 10-minute sensor streams, predict indoor air temperature, relative humidity, vapor pressure deficit, and yield.

Core story: VPD is the most operationally useful target—it governs crop metabolism and can drive climate control. NARX VPD R² reached 99.1% in the R greenhouse on a 70:15:15 split, with no meaningful gap between measured and predicted series.
2Adjacent single-span houses
NARXDynamic neural network
99.1%Peak VPD R² (RGH)
10 minSensor interval
9Predictor inputs
CC BY-NCOpen-access license
Mohammadreza Narimani greenhouse system diagram showing macroclimate disturbances, covering and HVAC inputs, and indoor temperature humidity VPD radiation and CO2 outputs
Figure 1. Protected-cultivation system: outdoor disturbances and indoor climate outputs.
Mohammadreza Narimani diagrams of R greenhouse and Q greenhouse thermal curtain, circulating fan, and side-vent layouts used in NARX climate modeling
Figure 2. Thermal-curtain layouts in the R greenhouse (RGH) and Q greenhouse (QGH).
Mohammadreza Narimani sensor layout map for two adjacent experimental greenhouses used in dynamic neural-network thermal modeling
Figure 3. Sensor locations in both experimental greenhouses.
Mohammadreza Narimani nonlinear autoregressive exogenous NARX architecture with input delay, feedback delay, weights, hidden layers, and output
Figure 4. Nonlinear autoregressive exogenous (NARX) model architecture.
Mohammadreza Narimani MATLAB NARX neural-network diagram with delayed x(t) and y(t) inputs, hidden layer of 10 neurons, and y(t) output
Figure 5. MATLAB NARX neural-network diagram used for greenhouse climate prediction.

Predicted Microclimate

Inputs included indoor air temperature, RH, VPD, inner-roof temperature and humidity, solar radiation, and outdoor temperature. Models were scored with R², MSE, RMSE, mean absolute deviation, and Nash–Sutcliffe efficiency. Predicted RGH means were about 20.1 °C, 50% RH, and 1.31 kPa VPD; QGH was slightly cooler and more humid. ANOVA found a significant VPD difference between houses at 5%.

Mohammadreza Narimani NARX predicted versus measured indoor air temperature time series for R greenhouse and Q greenhouse over eight December days
Figure 6. Predicted versus measured indoor air temperature in RGH (a) and QGH (b).
Mohammadreza Narimani NARX predicted versus measured relative humidity time series for R greenhouse and Q greenhouse
Figure 7. Predicted versus measured relative humidity in RGH (a) and QGH (b).

The authors emphasize VPD for control because NSE approached 1.00 for that target and because QGH spent more time in the agronomically preferred VPD band. The NARX VPD model reached R² of 99.1%, 98.7%, and 98.6% in RGH on 70:15:15 splits. Adjacent houses with different curtain geometry are therefore not interchangeable climate systems—even under the same outdoor weather.

Dynamic memory

NARX keeps lagged indoor states, which static regressions miss in greenhouse time series.

Curtain contrast

The only intended structural difference is screen geometry, isolating that design choice.

Control-ready VPD

High VPD fidelity supports using the model as a setpoint predictor, not only a hindcast.

Paper and Citation

Akpenpuun, T. D., Ogunlowo, Q. O., Na, W. H., Dutta, P., Rabiu, A., Adesanya, M. A., … & Lee, H. W. (2024). Dynamic neural network modeling of thermal environments of two adjacent single-span greenhouses with different thermal curtain positions. Journal of Agricultural Engineering, 55(2), 1563.

DOI: 10.4081/jae.2024.1563 · Open access under CC BY-NC 4.0

Please cite the publisher version when you use this work. The summary on this page is a research card, not a substitute for the article.

Cite and read the paper

Contact

Mohammadreza Narimani
PhD Candidate, University of California, Davis

📧 mnarimani@ucdavis.edu  |  🎓 Google Scholar