Greenhouse cultivation helps to create favorable microclimates where production
of vegetables and flowers is made possible through out the year or part of the
year as per the requirement. The greenhouse technology has also tremendous scope
in horticultural sector, especially for production of hybrid seeds, high value
vegetables, ornamental and medicinal plants (Kumar et
al., 2006). However, the quantity and quality in greenhouse crop production
can be realized only if the growth factors are managed so, they are available
to the plant at the right time with desired amounts (Portree,
1996). In fact, according to Hanan (1998), enriching
greenhouse just to maintain CO2 concentrations of 335 ppm (34 Pa)
can increase output and value of long season tomato crops by as much as 10%
with report in yield increases to 4 kg m-2. Regulating these environmental
factors to the benefit of the greenhouse crop is part of good greenhouse management.
The effect of temperature on tomato growth and its measurement is strongly linked
to climatic factors such as humidity, radiation, wind velocity and CO2
concentration. The last factor can limit growth because it is required for photosynthesis
(Nederhoff and Vegter, 1994).
Single control algorithms with separated control loops are not able to manage
the micro climate complexity. The interaction between microclimate factors and
plant responses needs data processing and a predicting tool of some variables
for process control purposes. A combination of new control methods based on
Artificial Intelligence (AI) technology with multi layer structures of control
systems was presented several years ago (Hashimoto, 1997;
Morimoto et al., 1996). Another strategy for
operating with predictive algorithms is the integration of micro climate models
into the control system shown by Schmidt (1996). This
in part reflects the fact that research is still revealing shortcomings in the
quantitative understanding of some processes.
Due to the fact that greenhouse system is considered as a nonlinear, multivariable,
nonstationary system and open to exterior environment (Bennis
et al., 2008), the black box models have been successfully used in
this kind of problems like Artificial Neural Networks (ANN) which according
to Chen et al. (2008) are useful when the rules
are unknown and it allows good fits to complex multivariate data.
Artificial Neural Networks (ANN) excel at uncovering patterns or relationships
in data and are also powerful non-linear estimators. Because of its empirical
nature, ANN technology is sometimes erroneously referred to as an advanced type
of regression analysis. However, regression treats all output variables independent
of each other, the presence of common arcs in the ANN architecture allows it
to identify important inter-relationships that may exist between output variables
Artificial Neural Networks (ANN) have been use widely in agriculture, examples
can be found by Patel et al. (1998), detecting
blood defects in eggs with a high precision of 91.1%. Also, Kalogirou
(2000) showed many ANN applications in energy problems showing that ANN
is a robust and tolerant method to white noise which is very convenient for
Seginer (1997) used Neural Network models for optimal
environmental control and he found that NN models relieves the difficulty involved
in finding a physical model which relates ventilation rate to outside conditions.
Linker et al. (1998) extended the use of Neural
Network Models by not only fitting models to experimental data, but also by
using them to optimize the greenhouse operation. A data based greenhouse model
was developed and incorporated in an optimization scheme to compute optimal
temperature and CO2 concentration set points as a function of the
changing weather. This study indicated that the models used not only fitted
the data well but also produced reasonable optimization results to control the
The final purpose of the climate control is to minimize the production cost in energy and raw materials but at the same time to have good quality products. In this particular study, we want to control temperature and CO2 enrichment inside of the greenhouse. Three models are developed, the first one is the temperature model, our concern is the temperature variations inside of the greenhouse, a good quality temperature sensors are still very sensitive to door openings and other factors so there is a need to eliminate this noise and to know the actual temperature trend in order to turn off/on the heating or cooling system, so the inside temperature is predicted 5 and 10 min ahead of the sensors reading.
The second and third model are linked, first we predict CO2 concentrations 5 min in advance to be used in the third model for the purpose of photosynthesis prediction. Given that CO2 injection is very expensive we do not want to apply CO2 when photosynthesis is decreasing. These two actions will result in money and energy savings.
MATERIALS AND METHODS
A Multi-Layered Perceptron (MLP) network with a hyperbolic function was used in this work for the three models mentioned above. The performance measure is the Mean Square Error (MSE) given by Eq. 1:
||Mean square error
||Total No. of data
||No. of output variables
The software package MATLAB Neural Networks Toolbox was used to train, validate and test the ANN prediction models.
This particular study was developed in a experimental venlo-style glass greenhouse (Cabin 6) at Humboldt University of Berlin. The greenhouse is 9.6 m width by 12 m long, 4.1 m height until ridge, the ratio of the roof vent opening compared to roof area is 50%, floor space is 115.2 m2 with 128 tomato plants (8 times 18). Sensors for all of the inside environmental variables were located in the center of the greenhouse at a height of 1.5 m for the period of study.
According to Kumari et al. (2006), the temperature
fluctuation inside of the greenhouse plays an important role for plant health.
Therefore, thermal heating inside of the greenhouse should be minimum, which
implies to maintain a good temperature control close to the optimum levels.
So, the set points for temperature where fix in 18°C to start the heating
system and 27°C to activate ventilation.
Data were collected every 5 min (this is the time step) from October 2007. The input variables for the temperature model were: outside average air temperature (Tex, °C), outside relative humidity (RHex, %), solar radiation (SR, W m-2), wind velocity (WV, m sec-1), inside relative humidity (RH, %), inside air temperature 5 min lagged and actual temperature (Tt-1, Tt, °C), vents opening left (Vleft, %) and right (Vright, %), average temperature in the heating system in the top (HStop, °C) and the bottom (HSbottom, °C) calculated as the average between temperature of the water in the inlet and outlet in the heating system. The variables used in the output layer were inside air temperature measured 5 and 10 min ago (Tt+1, Tt+2,°C).
A total number of 7800 data points were available, of which 50% were used for
training, 25% for verification and 25% for testing (Swingler,
2001). During training, data patterns were processed through the ANN until
a minimum acceptable error between measured and predicted values in Eq.
1 was achieved. It was at this point that the ANN had learned to predict
the system behavior of interest (i.e., values of output variables) in response
to the values of the input variables.
As shown in Fig. 1a-e, for some periods
where there are good conditions for photosynthesis (temperature, PAR), there
is not enough CO2 supply (red arrows) to increase photosynthesis,
so it is necessary to improve the CO2 control by having a good prediction
of both variables, in this way, we can stop the CO2 injection when
Variables behavior inside of cabin 2 June 2008
Leaks in the greenhouse allow a continuous infiltration of outside air. An
average value for infiltration in a glass house would be one air change h-1
(Blom et al., 2002); therefore outside CO2
concentration (CO2ext, ppm) and wind velocity (WV, m sec-1)
are included as an input variables in the model.
Data was recorded every 5 min from June 2007 for Cabin 2. This Cabin have the same characteristics as Cabin 6, but with CO2 enrichment. The experimental greenhouse has three pipes on the floor for CO2 enrichment, each pipe is 10 m long and they are uniformly distributed, the CO2 valve is open every time the CO2 level is below 250 ppm in which case a gas rate of 10 lt min-1 is released. The CO2 valve is closed when CO2 level inside of the greenhouse is 1000 ppm.
The variables used in the input layer for the CO2 model are: outside average air temperature (Tex, °C), outside CO2 concentration (CO2ext, ppm), solar radiation (SR, W m-2), wind velocity (WV, m sec-1), inside temperature (T, °C), inside CO2 (CO2, ppm) and inside relative humidity (RH, %), vents opening left (Vleft, %) and right (Vright, %). The output layer has only one variable: inside CO2 concentration 5 min ahead (t+1), 50% of the data was used for training, 25% for verification and 25% for testing. An additional data set of 200 values were used to test the performance of the Neural Network model. For this particular set of data the outside CO2 concentration varies from 306-498 ppm while the inside CO2 concentration varies from 283-1097 ppm.
Carbon dioxide is an important factor for the photosynthesis process, it
enters into the plant through the stomata openings by the process of diffusion.
The stomata represents the prime passageways for the escaping water vapour.
At the same time the stomata are the entrance openings for CO2. By
closing their stomata, plants are in a position to cut down the water loss,
however stomata closure has a harmful consequence for photosynthesis as the
CO2 supply is also cut (Ehlers and Goss, 2003).
To account for the stomata openings we measure stomata conductance at the leaf
surface under natural conditions, a cuvette is clamped to the leaf and the environmental
factors are recorded and mimicked inside the cuvette.
For this particular study, the device for photosynthesis measurement is the PTM-48M photosynthesis, an automatic four channel open type system for monitoring CO2 exchange and transpiration leaves. Leave cuvetts are fix at the first full developed leave at the top for measuring photosynthesis. In the open photosynthesis system the CO2 exchange is determined on the basis of the depression of CO2 concentration at the outlet (CO2out) of the leaf chamber in comparison with the incoming ambient air (CO2in). The CO2 exchange rate is calculated in Eq. 2:
||CO2 exchange rate
||Air flow rate through leaf cuvette (m3 h-1)
||Dimension factor (depends on pressure and temperature and is calculated
Photosynthesis is calculated taking into account the air flow rate through leaf cuvette (m3 h-1) and the CO2 molar mass and density.
The output from CO2 model was used as an input in the photosynthesis model. However, given that photosynthesis is measured every 15 min we applied linear interpolation to generate two additional points, in this way we can have a point every 5 min as the rest of the data in order to fulfill uniformity. As in the CO2 model, data from Cabin 2 were used every 5 min from June 2007 at the experimental Venlo-style glass greenhouse at the Humboldt University of Berlin. The variables used in the input layer are: inside air temperature (Tt, °C), inside CO2 concentration (CO2t, ppm), inside relative humidity (RHt, %), stomata conductance (STcond), Photosynthetically Active Radiation (PARt, μmol m-2*s), actual and 5 min lagged photosynthesis (Photot, Photot-1). The output in the model is the photosynthesis predicted 5 and 10 min ahead of time (Photot+1, Photot+2).
The Neural Network model for temperature prediction was trained, validated
and tested and the algorithm converges after 103 iterations with a low mean
square error of 0.0025. The optimal ANN architecture has 3 layers, the input
layer with 11 nodes (input variables), the hidden layer with 9 nodes and the
output layer with 2 nodes (prediction variables). A linear regression was performed
between measured and predicted temperature values for times 5 and 10 min ahead
(t+1 and t+2) and the correlation coefficients were 0.997 and 0.994, respectively.
This model was able to predict temperature 5 and 10 min ahead of the sensor
b) Temperature predictions using a different data set. NN predictions
cabin 6 October 2007
predictions using a different data set. One day CO2 predictions
cabin 2 June 22, 2008, 3:20-19:55 h (5-100 min)
In order to test the performance of the model a new data set was feed into
the ANN model, only the input data was provided. Figure 2a
and b show the simulation results, the MSE´s between measured
and predicted temperature values were 0.008 and 0.0295 for the temperature 5
and 10 min ahead, respectively.
For the CO2 model the algorithm converges after 82 iterations with a low mean square error of 0.00332. Again a linear regression was performed between measured and predicted CO2 values 5 min ahead (t+1) with a correlation coefficient of 0.9845. The ANN was feed with an additional 200 input data without providing the output variables, Figure 3 shows the results, the MSE´s between measured and predicted CO2 values was 535, which is a big value compared with the results obtained for temperature model. Sometimes, the measured values increase or decrease so rapidly while the predictions from the model are smooth, for these particular case there is a single row in the data set which make a big contribution to the MSE error (Fig. 3).
The predicted CO2 values were used as an input for the photosynthesis
model together with the other input variables. The ANN model for photosynthesis
prediction converges after 82 iterations with an MSE error of 0.004922. The
performance of the ANN model was tested with 150 different data set, for two
cases: for measured and predicted CO2 values as an input data, the
first case is shown in Fig. 4a and b and
the second case in Fig. 5a and b. There
is a small difference between the two cases but the MSE`s errors decrease when
the predicted CO2 values are used as an input data. As for the other
models a linear regression was performed between measured and predicted values
for the two cases resulting in correlation coefficients slightly bigger when
measured CO2 values were used instead of the predictions from NN
model (Fig. 4, 5a and b).
b) Photosynthesis predictions using predicted CO2 from ANN
model. Predictions using predicted CO2 cabin 2 June 6, 21:35
to June 7, 10:05 h, 2008
b) Photosynthesis predictions using measured CO2. Predictions
using measured CO2 cabin 2 June 6, 21:35 to June 7, 10:05 h,
Photosynthesis is a complex process which depends on so many variables specially
temperature (°C), CO2 (ppm), PAR (μmol m-2*s)
and relative humidity(%). The maximum photosynthesis (9.6 μmol m-2*s)
occurs when temperature (24.8°C), relative humidity (78.3%), PAR (663.55
μmol m-2*s) and CO2 (577 ppm) are in good levels.
For instance when PAR is maximum (1177.6 μmol m-2*s) and temperature
is 27.2°C, the photosynthesis is only 4.29 μmol m-2*s, because
the CO2 level is low (364 ppm). On the other hand, when CO2
level is maximum (1001 ppm) and relative humidity is high (86.1%), but PAR (323.15
μmol m-2*s) and temperature (22.6°C) are not in the optimal
levels, then photosynthesis is 8.09962 μmol m-2*s. A sensitivity
analysis was performed for the photosynthesis model (Table 1)
and the most important variables for the prediction were the photosynthesis
itself and relative humidity. Finally, Fig. 6 and 7
show the pathway for photosynthesis depending on relative humidity and temperature
and on relative humidity and PAR respectively, high levels of photosynthesis
are related with high levels of relative humidity.
pathway depending on temperature and relative humidity. Photosynthesis
pathway (June 10 from 4:50 am)
pathway depending on PAR and relative humidity. Photosynthesis pathway
(June 10 from 4:50 am)
analysis for photosynthesis model
open loop control with artificial neuronal networks
Three Neural Network models were successfully developed for temperature, CO2 and photosynthesis prediction in two experimental greenhouses located at the Humboldt University of Berlin. For the temperature model we use cabin 6 which is the control cabin naturally ventilated with tomato crop. The temperature predicting model was able to predict temperature 5 and 10 min ahead of the sensor signal with high precision.
A data set from cabin 2 with CO2 enrichment was used for the CO2 and photosynthesis prediction models for June 2007. The predictions of the CO2 model were acceptable, the measured and predicted values follows the same tendency, however the MSE error was not small (535) because of the big changes in the measured CO2 in one point, but in general the model provide accurate results.
On the other hand, the photosynthesis model was tested with measured and predicted CO2 values as an input data, in both cases the model provides highly precise results with a slightly differences, given the best results when predicted CO2 is used as an input variable. According to the sensitivity analysis relative humidity is an important variable in the photosynthesis process. Our final goal is to use the predicting models in an open loop control system (Fig. 8) and the ANN could both replace the sensors and give predictive information for a higher control quality. An essential prerequisite is an intermittent running of the training process so that the ANN can adapt to changings in the technical and biological constraints.
Authors would like to thank Conacyt (National Counselor in Science and Technology
in México) and DFG (German Research Foundation) for supporting the project
Environmental control and efficient energy use in tomato production under extreme
conditions in Germany and Mexico within Bilateral Cooperation Program. Also,
we appreciate the financial support from Agricultural Engineering and Water
Integral Use Graduate Program from Autonomous University of Chapingo, for the
publication of this article.