recognition. Intelligence is added to the process by computing the degree of
uncertainty, and computing with linguistic terms (fuzzy variables). More accuracy
is obtained compared to mechanistic models. Furthermore, the number of
parameters used to describe the system behavior is reduced.
Most of the studies implementing neural networks on greenhouses have focused
on modeling the dynamics of the greenhouse environment, targeting only air
temperature (Ferreira et al. 2002; Frausto and Pieters 2004; Linker and Seginer
2004; Lopez-Cruz et al. 2007; Patil et al. 2008; Seginer 1997) or both air temperature and relative humidity (He and Ma 2010; Salazar et al. 2008), or a combination of air temperature for CO 2 control (Linker et al. 1998; Salazar et al.
2010). Recent developments include hybrid systems based on neuro-fuzzy models
(Lopez-Cruz and Hernández-Larragoiti 2010; Qiuying et al. 2010; Yousefi et al.
2010) for modeling both air temperature and relative humidity. Recent techniques
have included modeling the greenhouse environment with hierarchical fuzzy
modeling (Salgado and Cunha 2005), or controlling the environment with optimized fuzzy control (Ehrlich et al. 1996; Lafont and Balmat 2002).
However, very few studies have focused on the development of climate controllers using neural networks. For example, Pucheta el al. (2006) implemented a
neurocontroller based on Neuro-dynamic programming on a lab-scale greenhouse
for the optimal production of tomato seedlings. The neurocontroller drives the
microclimate that controls the growth and development of the crop, while minimizing the operative costs and the final state errors on process variables and
actuator signals. Fourati and Chtorurou (2007) implemented an Elman neural
network to emulate the direct dynamics of the greenhouse, and trained a multilayer
feedforward neural network (MLFFNN) to learn the inverse dynamics of the
control process. The latter was placed in cascade with the Elman network to drive
the system output to desired values. In a similar way Fitz-Rodríguez et al. (2012)
proposed a Neural Network Predictive Controller to control the greenhouse climate of a naturally ventilated and fog-cooled greenhouse. Under this architecture
an MLFFNN was trained to predict the system dynamics. The system controls
fogging rates (given by pump pressure and the ON-time) and vent openings to
drive the inside climate to the set points while achieving a uniform environment.
Fewer studies implement control strategies that model the whole biosystem,
focusing either on the environment or on the crop component. In this way Tien
(1997) implemented a hybrid Neuro-Fuzzy approach for the system identification
and modeling of the total dry weight yield of tomato and lettuce. Center and
Verma (1997) developed a fuzzy model to predict net photosynthesis of tomato
crop canopies, and the results were highly correlated to the outputs of the mechanistic model TOMGRO to model crop processes. Ehret et al. (2008), used artificial neural networks to predict cuticle cracking in greenhouse peppers and
tomatoes. Fitz-Rodríguez and Giacomelli (2009) implemented a dynamic neural
network for modeling and predicting weekly production rates of greenhouse
tomatoes as a function of environmental parameters and current and past crop
conditions.
414
I. L. López-Cruz et al.
uncertainty, and computing with linguistic terms (fuzzy variables). More accuracy
is obtained compared to mechanistic models. Furthermore, the number of
parameters used to describe the system behavior is reduced.
Most of the studies implementing neural networks on greenhouses have focused
on modeling the dynamics of the greenhouse environment, targeting only air
temperature (Ferreira et al. 2002; Frausto and Pieters 2004; Linker and Seginer
2004; Lopez-Cruz et al. 2007; Patil et al. 2008; Seginer 1997) or both air temperature and relative humidity (He and Ma 2010; Salazar et al. 2008), or a combination of air temperature for CO 2 control (Linker et al. 1998; Salazar et al.
2010). Recent developments include hybrid systems based on neuro-fuzzy models
(Lopez-Cruz and Hernández-Larragoiti 2010; Qiuying et al. 2010; Yousefi et al.
2010) for modeling both air temperature and relative humidity. Recent techniques
have included modeling the greenhouse environment with hierarchical fuzzy
modeling (Salgado and Cunha 2005), or controlling the environment with optimized fuzzy control (Ehrlich et al. 1996; Lafont and Balmat 2002).
However, very few studies have focused on the development of climate controllers using neural networks. For example, Pucheta el al. (2006) implemented a
neurocontroller based on Neuro-dynamic programming on a lab-scale greenhouse
for the optimal production of tomato seedlings. The neurocontroller drives the
microclimate that controls the growth and development of the crop, while minimizing the operative costs and the final state errors on process variables and
actuator signals. Fourati and Chtorurou (2007) implemented an Elman neural
network to emulate the direct dynamics of the greenhouse, and trained a multilayer
feedforward neural network (MLFFNN) to learn the inverse dynamics of the
control process. The latter was placed in cascade with the Elman network to drive
the system output to desired values. In a similar way Fitz-Rodríguez et al. (2012)
proposed a Neural Network Predictive Controller to control the greenhouse climate of a naturally ventilated and fog-cooled greenhouse. Under this architecture
an MLFFNN was trained to predict the system dynamics. The system controls
fogging rates (given by pump pressure and the ON-time) and vent openings to
drive the inside climate to the set points while achieving a uniform environment.
Fewer studies implement control strategies that model the whole biosystem,
focusing either on the environment or on the crop component. In this way Tien
(1997) implemented a hybrid Neuro-Fuzzy approach for the system identification
and modeling of the total dry weight yield of tomato and lettuce. Center and
Verma (1997) developed a fuzzy model to predict net photosynthesis of tomato
crop canopies, and the results were highly correlated to the outputs of the mechanistic model TOMGRO to model crop processes. Ehret et al. (2008), used artificial neural networks to predict cuticle cracking in greenhouse peppers and
tomatoes. Fitz-Rodríguez and Giacomelli (2009) implemented a dynamic neural
network for modeling and predicting weekly production rates of greenhouse
tomatoes as a function of environmental parameters and current and past crop
conditions.
414
I. L. López-Cruz et al.