Multi-layer Perceptron Neural Network to Assess the Thermal

Behaviour of a Moroccan Agriculture Greenhouse

Noureddine Choab

, Hamza Ali-Ou-Salah

, Abdeljalil Jikal

, Said Saadeddine

and Ataa Abouatallah

Laboratory of Atmosphere's Physics, Materials and Modeling, Hassan II Casablanca University, BP 146, Mohammedia,

Morocco

Laboratory of Physic of Condensed Matter & Renewable Energy, Hassan II Casablanca University, BP 146,

Mohammedia, Morocco

Laboratory of Applied Chemistry and Environment, National School of Applied Science , IbnZohr University, PO Box

1136, 80000 Agadir, Morocco

Keywords: Machine learning, Neural networks, greenhouse thermal behaviour, prediction.

Abstract: The aim of this research is to demonstrate how machine learning algorithms can estimate the temperature

and relative humidity of a Moroccan greenhouse's inside air. The prediction model was created using a

Multi-Layer Perceptron neural network (MLPNN) trained using the Levenberg-Marquardt backpropagation

technique. The weather data and indoor air temperature and relative humidity of a greenhouse located in

Agadir, Morocco were used. The results reveal that the MLPNN's Root Mean Square Error (RMSE) values

are very low, and the Karl Pearson's coefficient of correlation (R) values are very close to 1, indicating that

the MLPNN has a high predicting accuracy. In addition the result of the comparison between the results

obtained by the MLPNN model and the data from the experiment shows that the predicted and measured

indoor thermal behaviour are similar, which mean that the MLPNN have a high ability to predict the

greenhouse thermal behaviour.

1 INTRODUCTION

With the growing global demand for food, farming

in a controlled environment is an effective way to

produce plants year-round. The greenhouse system

is one of the main types of controlled agricultural

environments (Iddio et al., 2020). The greenhouse is

a translucent building that produces a microclimate

ideal for plants and shields them from the outside

environment by reflecting incident solar energy

(Choab et al., 2020). This helps to increase the

production and quality of these plants (Choab et al.,

2019).

The greenhouse is not entirely insulated from the

outside environment. Therefore, the thermal

behavior of indoor greenhouse air is strongly

affected by the outdoor climate. As a result, it's

difficult to control and predict the temperature and

humidity inside the greenhouse (Moon et al., 2018).

Accurate prediction of the greenhouse indoor air

temperature and humidity has been studied in

various studies (Yu et al., 2016).

Artificial neural network (ANN) is used for

greenhouse thermal modeling since as it is suitable

for nonlinear system models (Castañeda-Miranda

and Castaño-Meneses, 2020; Dariouchy et al., 2009;

Giuseppina Nicolosi et al., 2017; Wang et al., 2009).

Taki et al. (2016) compared the ANN technique to a

mathematical model to see which method was best

for forecasting the temperature of the inside air, the

temperature of the roof, and the energy loss in a

semi-solar greenhouse. This comparison revealed

that the ANN technique is a viable way for solving

the non-linear relationship between greenhouse

internal environment factors and forecasting interior

air temperature, cover temperature, and greenhouse

energy loss with high precision. The ANN was used

by Francik and Kurpaska (2020) to develop a model

that predicted temperature change within a heated

foil tunnel. Trejo-Perea et al. (2009) used an

MLPNN to predict greenhouse energy usage. The

temperature and relative humidity outputs are used

as inputs in the model's cascade architecture. When

compared to the regression model, Duncan's

Choab, N., Ali-Ou-Salah, H., Jikal, A., Saadeddine, S. and Abouatallah, A.

Multi-layer Perceptron Neural Network to Assess the Thermal Behaviour of a Moroccan Agriculture Greenhouse.

DOI: 10.5220/0010727400003101

In Proceedings of the 2nd International Conference on Big Data, Modelling and Machine Learning (BML 2021), pages 10-14

ISBN: 978-989-758-559-3

Multiple Range Test demonstrates that ANN has a

high confidence level of 95%. Yue et al. (2018) used

an upgraded Levenberg-Marquardt Radial Basis

Function Neural Network (LM-RBF) to construct a

model to predict the inside air temperature and

humidity of a greenhouse, and this model achieves

the goal with a maximum relative error of less than

0.5 percent.

From the previous literature review, it appears

that machine learning techniques show great

potential in predicting thermal behavior in

greenhouse applications. Also, the majority of

applications employed ANN as a prediction

technique. The aim of this research is to forecast the

internal air temperature and relative humidity of a

realistic greenhouse used for tomato cultivation in

Agadir, Morocco's semi-arid region.

2 MATERIALS AND METHODS

2.1 Greenhouse Description

The experiments were carried out in a Canarian

greenhouse in the Khmiss Ait Amira Souss Massa

region by APEF&V (Association of producers and

exporters of fruits and vegetables) (Figure 1). The

surface of the greenhouse’s ground is 8200 m²

(88.23 m length, 95 m width, a height of 5 m at

gutter level and 6 m at span level). Tomatoes are the

crop grown in greenhouses.

Figure 1: Schematic view of the Canarian-type greenhouse

Inside air temperature, outside solar radiation,

wind speed, ambient temperature, and relative

humidity are all needed in the greenhouse's air

temperature and relative humidity forecast model.

These data were collected using a weather station

located inside (middle) and outside of the

greenhouse. A number of 4000 values were obtained

as the main dataset. To measure the air temperature

and relative humidity inside and outside the

greenhouse, ADCON TR1 was used. The ADCON

Silicium-Pyranometer SP-Lite was used to measure

the outside solar irradiance. The wind velocity

measurement was provided by ADCON Wind

Sensor Set Pro 10/2. Table 1 shows the details of the

sensors used for the measurement.

Table 1: sensors details

Sensor

Parameter

measure

Accuracy

pt1000

(ADCON

TR1

)

air

temperature

±0.1% at 20 °C

HC101

(ADCON

TR1)

air relative

humidity

± 1% from 0 to

90% and ± 2%

from 90 to 100%

at 20 °C

ADCON

Silicium-

Pyranometer

SP-Lite

solar

irradiance

sensitivity

(nominal)

between 60 and

100 µV/W/m²

ADCON

Wind Sensor

Set Pro 10/2

wind

velocity

± 0.3 m/s

2.2 Artificial Neural Network

The artificial neural network (ANN) appeared as a

modeling and prediction technology due to its

flexible mathematical structure able to describe

complex non-linear relationships between inputs and

outputs (Castañeda-Miranda and Castaño, 2017).

The multilayer perceptron neural network (MLPNN)

is one of the ANN variants used (Ali-Ou-Salah et al.,

2021). In this research, MLPNN was developed to

accurately forecast the indoor air temperature and

relative humidity in a greenhouse. The MLPNN is

able to obtain an efficient approximation of

nonlinear functions by using three interconnected

layers. The first one is the input layer, the second

one consists of one or more hidden layers, while the

third one is the output layer (Ali-Ou-Salah et al.,

2021; Bahani et al., 2020) (Figure 2). Each layer is

made up of neurons, which are processing units. In

addition, each hidden and output layer neuron has an

activation function that determines its output.

Weights and biases are synaptic connections that

carry information from one layer to the next (Ali-

Ou-Salah et al., 2021). A backpropagation (BP)

training algorithm is used to modify weights and

biases in order to learn the network. The Levenberg

Marquardt (LM) algorithm was employed as a

training algorithm in this work.

Multi-layer Perceptron Neural Network to Assess the Thermal Behaviour of a Moroccan Agriculture Greenhouse

Figure 2: Structure of a multilayer perceptron feedforward

neural network.

The output of the artificial neuron is described by

Equation (1):

𝑦



𝑓

∑

𝜔



𝑥







𝑏



 (1)

where 𝑦



is the neuron output of the 𝑗-th hidden

layer, 𝜔



is the synaptic weight between the 𝑗-th

layer and the 𝑖-th layer, 𝑥



is the input of the 𝑗-th

layer, 𝑏



is the bias weight of neuron j and 𝑓𝑥 is

the activation function.

In the current study, the linear and hyperbolic

tangent functions were employed (Table 2). The

output layer typically uses linear functions, but the

hidden and output layers can apply non-linear

activation functions. The most used non-linear

activation functions is hyperbolic tangent, because it

is compatible with the back-propagation algorithm

(Escamilla-García et al., 2020).

Table 2: The used Activation functions in this study.

Name Graphic Function

Hyperbolic

tangent

𝑓



𝜉









∙

1,

interval [-1,1]

Linear

𝑓



𝜉



𝑎∙𝜉𝑏

The predictors are the available external weather

data that are recorded in the experimental setup of

the greenhouse. These data cover the ambient air

temperature, solar radiation, relative humidity and

wind speed. These variables are physically

influencing the thermal behavior of the indoor

environment of the greenhouse as highlighted by

many physical models as shown in (Choab et al.,

2019). These variables can be considered as

significant for predicting the internal air temperature

of the greenhouse.

2.2.1 The Architecture of MLPNN Model

In order to find the optimal architecture of the ANN

model, the grid search technique is applied using 5

folds cross validation method.

The grid search method involves investigating a

large number of hidden neurons in order to discover

the best design. 5 folds cross validation is used to

evaluate the constructed model, which consists of

randomly dividing the data set into five folds of

data, each of which is utilized as a testing set and the

rest as a training set. The average of all testing errors

is the 5 folds cross validation error. The grid search

technique's outcomes are shown in Figure 3.

Figure 3: The grid search technique for finding the optimal

number of hidden neurons.

The results shows that 30 neurons give the

lowest 5 folds cross validation error. As a

conclusion, the optimal architecture of the ANN

model is shown in the Figure 4.

Figure 4: the ANN architecture used for the current study

2.2.2 Model Evaluation Criteria

Two metrics indicators were applied to evaluate the

performance of the developed models (Alghamdi et

al., 2020):

 Karl Pearson’s Coefficient of Correlation

(R)

𝑅



𝐴,𝐵





∑



















∑























(2)

𝐴

and 𝐵



are the mean of 𝐴 and 𝐵 variables,

respectively.

 Root Mean Square error (RMSE)

𝑅𝑀𝑆𝐸 







∑

𝑌



𝑌













(3)

4,05

4,1

4,15

4,2

4,25

4,3

4,35

0 50 100

5foldscrossvalidation

error

Numberofhiddenneurons

BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)

𝑌



and 𝑌





are the actual and predicted values

values, respectively, whereas 𝑌



is the mean of

𝑌



and N represents the number of observations.

3 RESULTS AND DISCUSSION

The internal air temperature and relative humidity of

the greenhouse were predicted using an MLPNN

model trained using a Levenberg-Marquardt

backpropagation approach.

3.1 Prediction Performances

The results of the predicting stage were summarized

in figures 5 and in Table 3. In the regression plot

between the forecasted and measured values, it

could be noticed that the majority of data points are

falling on the regression line, which means that the

MLPNN model gives very accurate forecasting for

the greenhouse thermal behavior. The RMSE values

for the MLPNN, as shown in Table 3, are very low,

and the R values are very close to 1, which mean

that MLPNN have a high accuracy of forecasting.

Figure 5: The Regression plot between the

forecasted and measured values

Table 3: Performance metrics of the MLPNN model

Training Validation Testing

RMSE 4.686367 5.152746 4.637304

R 0.989980 0.987724 0.990269

3.2 Greenhouse Thermal Behavior

In this section, we randomly selected days from an

unused dataset (29 Avril to 3 Mai 2019), in order to

make a comparison between the results obtained by

the MLPNN model and the data from the

experiment. Based on the current model, the inside

air temperature and relative humidity profiles of the

greenhouse were obtained (Figure 6 and 7). It can be

seen that the predicted and measured indoor thermal

behavior are similar, which mean that the MLPNN

have a hight ability to predict the greenhouse

thermal behaviour.



Figure 6: Experimental and estimated indoor air

temperature of the greenhouse



Figure 7: Experimental and estimated indoor air relative

humidity of the greenhouse

4 CONCLUSION

The present work examined the potential of adopting

machine learning-based techniques for the prediction

20 40 60 80 100

Target

100

Output ~= 0.98*Target + 0.84

Training: R=0.98998

Data

Fit

Y = T

20 40 60 80 100

Target

100

Output ~= 0.98*Target + 0.97

Validation: R=0.98772

Data

Fit

Y = T

20 40 60 80 100

Target

100

Output ~= 0.98*Target + 0.77

Test: R=0.99027

Data

Fit

Y = T

20 40 60 80 100

Target

100

Output ~= 0.98*Target + 0.85

All: R=0.98969

Data

Fit

Y = T

Air Temperature (°C)

Air Relative Humidity (%)

Multi-layer Perceptron Neural Network to Assess the Thermal Behaviour of a Moroccan Agriculture Greenhouse

of air temperature and relative humidity inside a

Moroccan greenhouse for Tomato cultivation. The

main predictors were the available recorded external

weather data that can be easily obtained by low-cost

measurements, such as the ambient temperature,

solar radiation, relative humidity and wind speed.

The results shows that the RMSE values for the

MLPNN, are very low, and the R values are very

close to 1, which mean that MLPNN have a high

accuracy of forecasting. In addition the result of the

comparison between the results obtained by the

MLPNN model and the data from the experiment

shows that the predicted and measured indoor

thermal behavior are similar, which mean that the

MLPNN have a high ability to predict the

greenhouse thermal behaviour.

REFERENCES

Alghamdi, A. S., Polat, K., Alghoson, A., Alshdadi, A. A.,

& Abd El-Latif, A. A. (2020). A novel blood pressure

estimation method based on the classification of

oscillometric waveforms using machine-learning

methods. Applied Acoustics, 164, 107279.

Ali-Ou-Salah, H., Oukarfi, B., Bahani, K., & Moujabbir,

M. (2021). A New Hybrid Model for Hourly Solar

Radiation Forecasting Using Daily Classification

Technique and Machine Learning Algorithms.

Mathematical Problems in Engineering, 2021.

Bahani, K., Ali-Ou-Salah, H., Moujabbir, M., Oukarfi, B.,

& Ramdani, M. (2020, September). A Novel

Interpretable Model for Solar Radiation Prediction

based on Adaptive Fuzzy Clustering and Linguistic

Hedges. In Proceedings of the 13th International

Conference on Intelligent Systems: Theories and

Applications (pp. 1-6).

Castañeda-Miranda, A., & Castaño, V. M. (2017). Smart

frost control in greenhouses by neural networks

models. Computers and Electronics in Agriculture,

137, 102-114.

Castañeda-Miranda, A., & Castaño-Meneses, V. M.

(2020). Smart frost measurement for anti-disaster

intelligent control in greenhouses via embedding IoT

and hybrid AI methods. Measurement, 164, 108043.

Choab, N., Allouhi, A., El Maakoul, A., Kousksou, T.,

Saadeddine, S., & Jamil, A. (2019). Review on

greenhouse microclimate and application: Design

parameters, thermal modeling and simulation, climate

controlling technologies. Solar Energy, 191, 109-137.

Choab, N., Allouhi, A., El Maakoul, A., Kousksou, T.,

Saadeddine, S., & Jamil, A. (2020). Effect of

Greenhouse Design Parameters on the Heating and

Cooling Requirement of Greenhouses in Moroccan

Climatic Conditions. IEEE Access.

Dariouchy, A., Aassif, E., Lekouch, K., Bouirden, L., &

Maze, G. (2009). Prediction of the intern parameters

tomato greenhouse in a semi-arid area using a time-

series model of artificial neural networks.

Measurement, 42(3), 456-463.

Escamilla-Garcia, A., Soto-Zarazúa, G. M., Toledano-

Ayala, M., Rivas-Araiza, E., & Gastélum-Barrios, A.

(2020). Applications of Artificial Neural Networks in

Greenhouse Technology and Overview for Smart

Agriculture Development. Applied Sciences, 10(11),

3835.

Francik, S., & Kurpaska, S. (2020). The use of artificial

neural networks for forecasting of air temperature

inside a heated foil tunnel. Sensors, 20(3), 652.

Nicolosi, G., Volpe, R., & Messineo, A. (2017). An

innovative adaptive control system to regulate

microclimatic conditions in a greenhouse. Energies,

10(5), 722.

Iddio, E., Wang, L., Thomas, Y., McMorrow, G., &

Denzer, A. (2020). Energy efficient operation and

modeling for greenhouses: A literature review.

Renewable and Sustainable Energy Reviews, 117,

109480.

Moon, T. W., Jung, D. H., Chang, S. H., & Son, J. E.

(2018). Estimation of greenhouse CO 2 concentration

via an artificial neural network that uses

environmental factors. Horticulture, Environment, and

Biotechnology, 59(1), 45-50.

Taki, M., Ajabshirchi, Y., Ranjbar, S. F., Rohani, A., &

Matloobi, M. (2016). Heat transfer and MLP neural

network models to predict inside environment

variables and energy lost in a semi-solar greenhouse.

Energy and Buildings, 110, 314-329.

Trejo-Perea, M., Herrera-Ruiz, G., Rios-Moreno, J.,

Miranda, R. C., & Rivasaraiza, E. (2009). Greenhouse

energy consumption prediction using neural networks

models. training, 1(1), 2.

Wang, D., Wang, M., & Qiao, X. (2009). Support vector

machines regression and modeling of greenhouse

environment. Computers and electronics in

agriculture, 66(1), 46-52.

Yu, H., Chen, Y., Hassan, S. G., & Li, D. (2016).

Prediction of the temperature in a Chinese solar

greenhouse based on LSSVM optimized by improved

PSO. Computers and Electronics in Agriculture, 122,

94-102.

Yue, Y., Quan, J., Zhao, H., & Wang, H. (2018, August).

The Prediction of Greenhouse Temperature and

Humidity Based on LM-RBF Network. In 2018 IEEE

International Conference on Mechatronics and

Automation (ICMA) (pp. 1537-1541). IEEE.



BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)