Review. Advantages and disadvantages of control theories applied in greenhouse climate control systems

Today agriculture is changing in response to the requirements of modern society, where ensuring food supply through practices such as water conservation, reduction of agrochemicals and the required planted surface, which guarantees high quality crops are in demand. Greenhouses have proven to be a reliable solution to achieve these goals; however, a greenhouse as a means for protected agriculture has the potential to lead to serious problems. The most of these are related to the inside greenhouse climate conditions where controlling the temperature and relative humidity (RH) are the main objectives of engineering. Achieving appropriate climate conditions to ensure high yield and quality crops reducing energy consumption have been the objective of investigations for some time. Different schemes in control theories have been applied in this field to solve the aforementioned problems. Therefore, the objective of this paper is to present a review of different control techniques applied in protected agriculture to manage greenhouse climate conditions, presenting advantages and disadvantages of developed control platforms in order to suggest a design method-ology according to results obtained from different investigations.

In this review, updated information is provided about modern methods to control the greenhouse environments which can be taken into account as criteria to design new greenhouse microclimate control systems. The paper is divided in four sections, the first focuses on the different control theories applied to design climate control systems for protected agriculture. The second section is an overview of the technology platforms where the controllers were implemented. The third section discuses new tendencies in the development of environmental controllers for protected agriculture. Finally, in the last section the conclusions are presented.

Introduction
A greenhouse is an enclosed space that creates a different environment to that found outside due to the confinement of the air and to the absorption of shortwave solar radiation through a plastic or glass covers (El Ghoumari et al., 2005). This generates a new environment inside the greenhouse which is better known as microclimate. The greenhouse microclimate can be manipulated by control actions, such as heating, ventilation, CO 2 enrichment to name a few; in order to provide appropriate environmental conditions (Bennis et al., 2008). These modifications imply additional use of energy in the production process. Furthermore, it requires a control system that minimizes the energy consumption while keeping the state variables as close as possible to the optimum crop physiological reference (Coelho et al., 2005). Horticulture in greenhouse conditions is a rapidly expanding interest and is consequently increasing in its economic and social importance.
Many efforts have been made to develop advanced computerized greenhouse climate control systems. In particular, interesting and important optimal control approaches have been proposed (Ioslovich et al., 2009). As was previously reported, crop production using controlled environments has several advantages over conventional crop production such as greater productivity, better product quality, and low water and fertilizer consumption. Nevertheless, environmental requirements for living systems are very complex and nonlinear. Furthermore, the biological system likely has a significant and multiple effects on its physical surroundings (Pasgianos et al., 2003). Researchers have used many control techniques in different fields. From the conventional or sometimes referred to as classic Abbreviations used: AGA (annealing genetic algorithm); AI (artificial intelligence); ANN (artificial neural network); DIF (difference between average day temperature and average night temperature); DSP (digital signal processor); FL (feedback linearization); FLS (fuzzy logic system); FPGA (field programmable gate array);GA (genetic algorithm); MIMO (multiple-imput-multiple-output); MPC (model predictive control); PC (personal computer); PD (proportional derivative control); PDF (pseudo-derivative-feedback); PI (proportional integral control); PID (proportional integral and derivative control); PMP (pontryagin's maximum principle); PSO (particle swarm optimization); RH (relative humidity); SCS (sequential control search). Nomenclature: A g (covered ground surface, m -2 ); A r (roof to soil rate, m 2 m -2 ); C (greenhouse heat capacity, J K -1 m -2 ); C p (air specific heat, J K -1 kg -1 ); e i (internal mean vapour pressure, Pa); e(t) (error); E(s) (Laplace error representation); G (outside short-wave radiation, W m -2 ); I LA (leaf area index, m 2 m -2 ); K i (integral gain); K d (derivative gain); K out,air (heat loss coefficient from greenhouse air to outside air); K p (proportional gain); q h (heat input, W m -2 ); r s (stomatic resistance, s m -1 ); r a (aerodynamic resistance, s m -1 ); s (Laplace transform parameter); t (time, s); T c (crop temperature, K); T d (adjustment coefficient); T G (internal air temperature, K); T g (ground temperature, K); T i (adjustment coefficient); T o (external air temperature, K); T r (roof temperature, K); u(t) (process input); U(s) (Laplace input representation); V l (greenhouse air to soil area rate, m 3 m -2 ); x i (internal absolute humidity, kg m -1 ); x g (soil absolute humidity, kg m -3 ); x o (external absolute humidity, kg m -3 ); y(t) (process output); y d (t) (setpoint or desired process output); α ci (convection heat transfer coefficient, W m -2 ). Greek letters: γ (thermodynamic constant, Pa K); δ (leaf slope, Pa K); φ v (ventilation rate, m 3 s -1 ); η (radiation conversion factor); λ (water vaporization energy); ρ a (air density, kg m -3 ). Superscripts: * indicates that considered quantity is saturated at vapour pressure.
interest. This problem has been the focus of many researchers worldwide who have analyzed, experimented and proposed many climate control systems in order to manipulate variables such as temperature, relative humidity (RH), CO 2 enrichment, radiation and many others that are necessary to generate the fundamental conditions for successful protected agriculture.
Since most control theories require the mathematical model of the system for tuning and simulating the proposed algorithms, different greenhouse models have developed. It includes simple models that only describe air temperature to detailed models that even involve crop response. The traditional greenhouse climate models are based on energy and mass balances (Setiawan et al., 2000).
A model based on aforementioned balances over an elementary volume of greenhouse air was proposed by Arvanitis et al. (2000). Here, air temperature is represented by a differential Eq. [1]: where the T G is the greenhouse internal air temperature, C the greenhouse thermal capacity, K out, air is the heat loss coefficient from greenhouse air to outside air. T o is the external air temperature and q h is the heating power.
Recently, more detail models have been used for control proposes (Castañeda-Miranda et al., 2006). Those models involve almost all variables that influence the greenhouse behavior.
In Eqs.
[2] and [3], T c is the crop temperature, T g is the ground temperature, T r is the roof temperature, x i is the internal absolute humidity, x o is the external absolute humidity, x g is the soil absolute humidity, e i is the internal mean vapour pressure, C p is the air specific heat, V l is the greenhouse air to soil area rate, A g is the covered ground surface, A r is the roof to soil rate, r s is the stomatic resistance, r a is the aerodynamic resistance, G is the outside short-wave radiation, I LA is the leaf area index, α ci is the convection heat transfer coefficient, δ is the leaf slope, ρ a is the air density, λ is the water vaporization energy, η is the radiation conversion factor, γ is the thermodynamic constant, φ v is the ventilation rate and t is the time in s. The superscript * indicates that consider quantity is at saturated vapour pressure.
Finally, models that consider greenhouse-crop interaction and complex processes such as photosynthesis or transpiration have been developed (Van Straten et al., 2000). By this way, the greenhouse system can be represented concisely in a space state form as: Here, f{}s represent vector functions of the argument between brackets; x g , x b , x c are the state vector of the greenhouse (g), the storage buffers (non-structural biomas) (b) and the crop (c), respectively; u m , u a, the vectors of the manipulated control (m) and ambient (a) inputs, respectively; q ga , q gm , flux vectors representing fluxes between the greenhouse and the ambient environment (ga) and the operating equipment (gm), respectively; q gb , q gc , fluxes between greenhouse and buffer (gb) and between greenhouse environment and crop (gc), respectively; q cc , flux vectors related to the internal greenhouse conditions x g and the crop states x c , but not directly on the ambient conditions outside the greenhouse. The state list typically consist of temperature, moisture content and carbon dioxide for greenhouse atmosphere, temperatures for the storage buffers, and various crop biomass states for the plans in the greenhouse.
It is easy to find how models have been improved in response to production schemes that require more precisely methods to control the environment greenhouse.
In this section, an analysis and classification of the different control theories is presented. Establishing a division between the controllers presented in current literature is complicated, due to the variety and integration of diverse techniques used to solve the same problem. Fig. 1 shows a classification proposed dividing Review. Control systems applied in precision agriculture greenhouse climate control task in two main fields. The first one is usually called conventional control which consists on control theories which try to control the greenhouse environment just by reducing the deviation between set points of the interest variables and measured values to zero. As examples of conventional control there are ON/OFF, PID, other classical controllers and also paradigms of AI such as ANNs, FLS, GAs, among others. The other field is optimal control, in which the requirement is to consider aspects such as greenhouse behavior, actuator capabilities, energy consumption and mainly crop response as input parameters of the control process. Here, Expert systems and Model Predictive Control (MPC) are the most common techniques. However, aforementioned AI-based techniques can be also considered as optimal control when they consider input parameters such as crop responses among others.

Conventional greenhouse climate control
The most representative component of this theory is the PID (Ang et al., 2005); which is a feedback mechanism commonly used in industrial control systems (Fig. 2). Therefore, it is necessary to explain each component and action of the PID controller.
-Proportional (P) control: In certain cases having a smooth control and an error that is almost zero in the steady state is desired, where the proportional controller is suitable for this type of plant since that a proportional controller provides a control signal that is proportional to the error, that is, it returns its input multiplied by the proportional gain (K p ), thus, the control signal is given as: and the transfer function is obtained by means of the Laplace transform as: -Integral action: When an integral action is implemented, the integral of the error is added to the control signal. If the error signal is large, then the control signal increases quickly, but if the error signal is small then the control signal increases slowly. It is remarkable that if the error approaches zero then the controller output would remain constant. Due to this feature, integral action can be used when a constant load is present in the plant; even when no error is present, the controller will keep on providing an output signal for compensation.
-Proportional integral (PI) control: Since a proportional control is not capable of compensating a load in the plant without error, the integral action is neces-  sary. Integral action can compensate and provide a zero error; the PI controller is given as: where T i adjusts the integral action and K p adjusts both the integral and proportional actions. Its transfer function is given as: -Proportional derivative (PD) control: The goal of a derivative controller is to provide a signal proportional to the signal error change rate, causing derivative action to be present only when there is a change in the error signal. In other words, the derivative action introduces damping to the system. The PD controller is given as: and whose transfer function is: -Proportional integral derivative (PID) control. By combining the three different control actions a PID controller is obtained by: and whose transfer function is: The PID controller is the most complete controller available and the most resorted to since it provides a quick response, a control signal that tends to provide stability to the system and a minimum steady state error. The PID controller is an important control tool for industrial processes and only three gains have to be tuned (Ogata, 2003;Dorf & Bishop, 2005).
Although the PID is the most utilized controller in industry and it is widely accepted for agricultural applications, it is not the only solution for agricultural problems. Occasionally it is not a good choice due to the absence of a reliable mathematical model within the system. Taking this into account, a climate control system based in ON/OFF operation has been proposed when the mathematical model is unknown, and this complicates the tuning of the controllers. It manages the times when actuators are turned on, and prevents external climatic changes issues based on recorded data (Ali & Abdalla, 1993). The aforementioned technique was studied by Hooper & Davis (1988), who implemented a controller based in an algorithm that modifies the greenhouse heating setpoints depending on previously achieved temperatures. This technique has shown good performance managing deviations in the setpoints through soft changes. Hooper (1988) also presented an integral greenhouse climate control by applying a mixture of controllers, a PI controller applied in heating and ventilation, and ON/OFF control applied for irrigation, pH, electrical conductivity and nutrients management. However, Setiawan et al. (2000) reported that a Pseudo-Derivative-Feedback (PDF) control presents better performance than a PI for agricultural application, due to the fact that PDF controls have better load handling capability than PI controls. PDF control was Process or Plant Figure 2. Block diagram of general PID controller. Review. Control systems applied in precision agriculture better than PI for systems without time delay and significantly better for systems with time delay. It is easy to find current investigative work regarding the field of greenhouse microclimate control; however, reported results reveal that aforementioned techniques are not the most adequate to solve the problems inherent in greenhouses. The reason lies in the fact that the model of a greenhouse is very complex and it has many non-linearities; consequently, this has encouraged the development of new control techniques that do not require the greenhouse mathematical model (Sigrimis et al., 2002).
In the past, the application of new and advanced techniques for control was limited because of the limited computational power that was then available. Controllers based on FLS, ANNs or GAs could not be implemented in the former technological platforms due to their high complexity. These unconventional techniques based on soft computing and computational intelligence are now gaining popularity in the field of agriculture. Several soft-computing, such as ANNs and knowledge-based systems have been implemented with significant success (Soto-Zarazua et al., 2010).
FLS controllers are conceptually very simple; they consist of an input stage, a processing stage and an output stage. The first stage maps sensors and other inputs to the appropriate membership functions and truth values. The processing stage invokes the appropriate rules and generates a result for each. Finally, the output stage converts the combined results into a specific control output value. Furthermore, ANNs is a knowledge paradigm and automatic processing system that attempts to imitate how the nervous system of animals works. The principal advantage of this technique is that it does not require a model of the system. The system is composed of neurons with propagation, activation, and transfer functions that are interconnected among them in an effort to reduce the error to zero. GAs is a heuristic research that mimics the process of natural evolution. This heuristic is routinely utilized to generate solutions to optimization and search problems. GAs generate solutions to optimize problems using techniques inspired by natural evolution such as inheritance, mutation, selection and crossover.
The relatively new field of evolutionary computing has become increasingly popular in recent years due to the development of powerful and low-cost computational systems. Because AI based systems have been updated and improved, these techniques solve the main problem of classic controls that being the identification of the system which is commonly nonlinear (Caponetto et al., 2002).
The classical solutions proposed are generally based on the linearization of the process behavior regarding the operating points. Other research has been carried out on this technique of linearization, not only dealing with the operating points, but also by taking into consideration all the input-output space to obtain several local linear models. The major difficulty with this technique is the model transition. Indeed, many techniques of modeling and identification based on FLS are often used for these types of systems (Trabelsi et al., 2007). Some controllers base their operation on the aforementioned paradigm, as proposed by Castañeda-Miranda et al. (2006) who implemented a FLS on a field programmable gate array (FPGA) to control the temperature of the greenhouse microclimate or Kurata & Eguchi (1990) who applied this theory in crop management for protected agriculture.
Other systems which are classified in AI techniques are knowledge-based systems, as the one proposed by Gauthier & Guay (1990) which manages the climate control and production. This system supports dynamic optimization and continuous greenhouse monitoring. The proposed prototype was designed using objectoriented programming, obtaining good performance in the problem area. These expert-systems have proven to be a reliable alternative in greenhouse control applications. Jacobson et al. (1989) reported an expert system to control misting. This system was based on a strategy of an experienced grower. Gauthier (1992) reported changes to this scheme in a system that supports various types of digital process controllers as well as the creation and deployment of knowledge based control strategies with the goal of being able to intervene in a wide number of areas such as crop protection, climate control, crop nutrition, operational and strategic planning. This scheme received additional improvements such as changing the heuristic knowledge of the growers for data routinely collected in a commercial greenhouse .
Fuzzy systems achieved important results in the field of climate control for protected agriculture. Moreover, it is necessary to have reliable information of the system behavior, and that is not sufficient with this requirement a correct abstraction to create rules based in heuristic and empiric knowledge of the grower's experience is also necessary.
ANNs have proven their strengths and flexibility to adapt to non-linearities and unexpected parameters of the system. Their main disadvantage is that their proper training requires large multi-dimensional sets of data to reduce the risk of extrapolation and the uncertainty about their response to inputs which differ in relation to the training information. Therefore, minimizing the dimensionality of the problem, both input and state vectors become of paramount importance (Seginer, 1997).
GAs as FLS and ANNs offer the ability to control the system with a good performance without the requirement to base their operation on plant identification of the system. Although GAs represents a solution in the control of nonlinear systems, the computational requirements have limited its use in situ applications until recently. The final conclusion is that the use of AI based systems is justified in control problems where the system plant is highly nonlinear or when the model is not reliable or has not been identified.
Each control technique offers solutions for specific problems. Unfortunately, a specific controller that deals with the different characteristics and limitations presented in highly non-linear and complex systems of greenhouse microclimate has not yet been found. Consequently, hybrid models which combine different control schemes have begun to appear.
Combinations of classical control theory with AI and considerations of crop process have been demonstrated as promising. A demonstration of this was reported by Pinon et al. (2005); he proposed a scheme for greenhouse temperature control using the advantages of combining Feedback Linearization (FL) and standard linear MPC. The discussed hybrid control structure, MPC + FL, offers a reliable solution of nonlinear control problems, transforming a non-linear greenhouse system subject to input constraints, to an optimized problem for a linear greenhouse system.

Optimal control
The advantages of using optimal instead of conventional greenhouse climate control can be summarized as follows. An optimal control approach to greenhouse climate control fully exploits scientific quantitative knowledge concerning the greenhouse, the greenhouse equipment and the crop, captured all in a mathematical dynamic model that deals with the problem of maximizing the profit, achieving welfare of the crop through practices that minimizes production costs (Van Straten et al., 2011).
In this section, climate controllers based on complex algorithms are discussed. An analysis of MPC, real time controllers, robust and non-linear control, feedforward and systems that take into consideration decision support tools to gain efficient temperature integration are presented. Also, a survey of reported methods that considers morphological and physiological characteristics of the crops is presented.
The need of guarantee yield and quality of greenhouse crops has demanded stricter control of plant climate; previously, controllers were utilized for the sole purpose of adjusting the microclimate variables, but with recently increasing costs of energy, an appropriate controller that also considers the energy consumptions is necessary. One example of this was reported by Nielsen (1995) who presented a computer algorithm design to distribute the energy demands of greenhouses, reducing the peaks presented in the actuator in the day-to-night and night-to-day transitions. Other authors, proposed more complex systems (Arvanitis et al., 2000;Davis & Hooper, 2002). These methods operate by considering that greenhouse parameters vary with operating conditions and applied a new poleplacement scheme. It estimates the unknown parameters of the greenhouse on-line from sequential data of the greenhouse temperature and the heating power which is recursively updated to obtain a slightly soft control. Sigrimis & Rerras (1996) reported a controller based on a linear model structure to track and predict greenhouse behavior as a Multiple-Input-Multiple-Output (MIMO) system. This method takes into consideration disturbances like uncontrollable inputs. Climate controllers systems have been improved in order to consider external disturbances and have the ability to compensate for them; even they take into account plant responses such as crop growth. Reports regarding feedforward controllers indicate that they are reliable and achieve good performance for greenhouse heating (Jewett & Short, 1992;Takakura et al., 1994). On the other hand, real time systems have also been utilized in greenhouse applications, improving the systems to the point where operator intervention is only required to define the constraints of the heating setpoints.
Another real-time control algorithm for generating optimal heating setpoints was presented by Chalabi et al. (1996). This method adjusts greenhouse temperature setpoints over a period of time to achieve energy savings for a tomato crop, justifying these control actions by using the results of physiological stud-Review. Control systems applied in precision agriculture ies showing that for some crops it is sufficient to maintain an average temperature in a greenhouse over a given period (Hurd & Graves, 1983;De Koning, 1990). The algorithm is based on a model of greenhouse energy requirements and on a numerical method for optimization, where the optimal control problem is converted into a non-linear programming problem solved by sequential quadratic programming.
AI was also applied combined with crop process knowledge to generate paradigms applied in protected agriculture. Fitz-Rodriguez & Giacomelli (2009) made use of FLS combined with ANNs to propose a better control strategy for agriculture under greenhouse conditions taking into account models of crop growth. However, one disadvantage of ANNs is that, it requires getting a considerable set of data to train the net. This problem was addressed by Linker et al. (1998) using previously acquired data over a two-month period in a commercial greenhouse to train the net. The resulting model not only fit with data, it also seemed qualitatively correct and produced reasonable optimization results in a scheme of CO 2 enrichment control.
Hybrid controllers which fuse FLS and GAs were presented by Goggos & King (2000) in a research project that applies qualitative reasoning and evolutionary computing in the design of optimal set points and control strategies for greenhouses. This fusion of different intelligent control techniques was also applied in an intelligent environment control for plant production systems. The author used a decision-marking system based on ANNs and GAs in order to optimize plant growth under hydroponics conditions and also identified the response of plant growth to the nutrient concentration (Hashimoto et al., 2002).
One of the main problems with climate control is the greenhouse dynamic model which is highly nonlinear. Consequently, many researchers have applied simplified models for the non-linear problem, such as Ioslovich et al. (1996), who designed a controller based on a simplified model of the crop growth with constraints on the control signals. The objective of this optimization was to take into account the cost of energy used by heating and ventilation systems.
When it is necessary to control non-linear systems where the plant model is unknown, the use of feedbackfeedforward control is an alternative. In this configuration, unexpected events are considered in the control model and the controller attempts to reduce the error to zero no matter the disturbances. These system characteristics were used to design a non-linear controller for coupled air temperature and humidity (Albright et al., 2002). A feedback-feedforward combination was also used by Pasgianos et al. (2003) in a system linearization and decoupling of a greenhouse, maintaining control of ventilation/cooling and moisture. This technique also served to compensate for external disturbances. Finally, the controller was designed to consider the actuators capabilities and saturation setpoints.
In this section, applied strategies in climate controller systems with the objective of saving energy are considered. Nevertheless control techniques that consider physiological plant processes are also studied and described in this section.
Horticultural research has indicated that for the majority of plants, crop growth responds to long-term average temperatures rather than specific day and night temperature profiles (Langhans et al., 1980;Miller et al., 1985). Based on this research it has been suggested that the heating set-point can be adjusted to ensure that a desired average temperature over a given period can be achieved and thereby energy savings obtained. This knowledge was applied by Sigrimis & King (2000) in the design of a tool available to exploit the interaction between photosynthesis and growth according to the intuition of the grower because this interaction is not well known for most plants. The method is based on varying heating set-points using previously recorded information in order to achieve the desired average for any user-defined period. The proposed system does not require weather information and the grower can also set safety limits as the ultimate minimum and maximum temperatures permitted.
These control schemes were exploited by authors like Gauthier et al. (1995), who presented control strategies applied in heat, cool and dry greenhouse air, and also in the regulation of CO 2 , light and irrigation. Considering that plant processes vary with the day and vegetative state of the plant, they do not require strict control of the microclimate all the time. With this in mind, temperature integration systems for greenhouse cultivation were developed by Körner & Challa (2003a). The concept considers different crop processes, and a decoupled process with fast temperature response (e.g. photosynthesis or stress) from a process with a slow response time. The objective was to improve the temperature integration concept by introducing dynamic temperature constrains; these flexible boundaries depend on the underlying crop process while increasing the potential for energy saving in greenhouses.
A different approach was proposed with a decision support tool that assists in choosing the most appropriate climate according to the week of the year in order to obtain the optimal gains of sustainability and plant quality. The greenhouse climate and crop model are studied separately and jointly considering the effects of six different regimes with increasing degrees of freedom for various climate variables (Körner & Van Straten, 2008) which include: crop model, temperature integration, dynamic humidity control and negative DIF regimes (DIF = the difference between average day temperature and average night temperature, and therefore reduces the use of chemical growth regulators).
MPC is an advanced control technique applied in the field of protected agriculture. The objective was to predict the greenhouse variables behavior. Developments in MPC algorithms for greenhouse operation which takes into account weather predictions to generate new optimal control problems for each update of forecasted weather information as solved numerically by linear programming were also developed (Gutman et al., 1993). A contribution to this scheme was offered by Van Straten et al. (2000) where information about crop growth simplifies the design of greenhouse control strategies to obtain a truly economical control strategy. This approach leads to the concept of selecting processes by time response where the short-term effects like photosynthesis and evapo-transpiration are dealt with by an automated model-predictive optimal controller, while the long-term effects are left to the grower.
Aiming energy saving proposes, MPC algorithm has seen advances that take into account constraints in both manipulated and controlled variables, using on-line linearization for a real-time application. This proposal was applied in greenhouse temperature regulation, achieving good performance and energy savings. The results were compared with a PID solution. MPC based controllers solve the problems commonly presented in PID systems (El Ghoumari et al., 2005). Due to the advantages presented by MPC, different strategies have been applied to design optimal MPC controllers. The Particle Swarm Optimization (PSO) was applied to design a model-based predictive greenhouse air temperature controller subject to restrictions; the model employs data from the climate inside and outside the greenhouse, as well as the control inputs and controller outputs. The operation principle ensures set-point tracking and minimizes control efforts. The conclusions presented show better efficiency over GAs and sequential quadratic programming methods (Coelho et al., 2005).
MPC with GAs facilitating the incorporation of energy and water consumption to adjust non-linear models parameters have been suggested. The combination of MPC and GAs permits the control of the greenhouse microclimate while achieving energy and water savings (Blasco et al., 2007). GAs in annealing form (AGAs) has also been applied for calibrating classical controllers such as PID, where the AGAs play a role in the parameter identification, demonstrating advantages over traditional GAs like premature convergence and low computing efficiency that are required to implemented these (Fan & Zuo-hua, 2006). Feed-forward neural networks have been applied in conjunction with simple neural models to drive the system outputs to desired values (Fourati & Chtourou, 2007).
Robust control also contributed to systems in protected agriculture because of its ability to deal with uncertain parameters, disturbances or modeling errors. It was applied focusing on managing the high correlations between air temperature and hygrometry (Bennis et al., 2008). Different methods have been applied attempting to find a reliable optimal control solution for greenhouse environment, utilizing concepts from advance sequential control search (SCS) or Pontryagin's maximum principle (PMP) (Seginer & McClendon, 1992), to systems that consider the crop model and its effects on greenhouse behavior (Jones et al., 1990). The objective of these approaches was to include weather and greenhouse-crop characteristics such as ventilation and stomatal resistance in the control actions (Baptista et al., 2010). Humidity control regimes were also proposed by using information about the vegetative state of the plant (Körner & Challa, 2003b).
Authors worldwide have dedicated time and substantial effort to develop not only control systems, but they have also been working to create strategies to ensure reliable measurements through the use of filters and signal processing techniques to guarantee good performance of controller systems (Ibrahim & Sørensen, 2010).
One of the main objectives of the aforementioned developments is energy savings; however, all require knowledge of plant processes which is limited and usually empiric or heuristic. Consequently, advanced smart sensors are being developed in order to measure crop specific characteristics such as plant transpiration dynamics and photosynthesis in order to understand how physiological processes occur in the plants and how they affect and modify their surroundings (Millan-Almaraz et al., 2010). Review. Control systems applied in precision agriculture Throughout this review, special attention has been directed to demonstrate that not only the controller is necessary to guarantee appropriate microclimate conditions, but a fundamental part of the design is the use of reliable systems which take into consideration the importance of failure detection tools. By applying a hybrid of physical/neural network models with robust failure detection, failures are correctly detected and identified, leading to a significant reduction of losses caused by failures (Linker et al., 2000).

Technological platforms applied on climate control theories implementation
Technological platforms are also important when a system is being developed to solve specific problems; protected agriculture is not an exception to this rule. Agronomy field imposed hard operating conditions and found it is necessary to strictly consider these restrictions. Despite the fact that greenhouse climate is not a fast response system, robust platforms that guarantee uninterrupted operation are required, flexibility is crucial to improve changes in the system in order to solve emerging needs and lower cost is also necessary to ensure success of a newly emerging crop production industry (Fang & Zhen-xiao, 2008).
The first and most popular platform chosen for greenhouse control applications has been the personal computer (PC). The use of PCs in greenhouse operations has created possibilities to implement complex algorithms that were impossible to apply in the past (Fang & Zhen-xiao, 2008). Consequently, the integration of new task modules, sensor and communication devices becomes easier (Ali & Abdalla, 1993). Different conFigurations of PCs and networks have been proposed to achieve more efficient greenhouse management (Hooper, 1988).
Commercial climate control computers and proprietary data-logger were also applied (Nielsen, 1995;Linker et al., 1998). These systems offer solutions for protected agriculture problems. However, PCs are not the most appropriate platforms for heavy duty field applications, and are characterized to be a noisy and harsh environment with high humidity rates that are subject to constant temperature changes. Consequently, PCs are susceptible to failures and damage caused by the greenhouse harsh environment. Another consideration when discussing the use of PCs is the high cost to integrate PC networks or property systems. Other technological platforms should be proposed to ensure reliable sustainability. Microcontroller and Digital Signal Processors (DSP) based systems have attempted to solve the aforementioned problems with promising results; however, their limited capacities have proven to be difficult in the application of advanced algorithms with considerable computational demands (Coelho et al., 2005).
The development of embedded systems for a particular application has been demonstrated to be the best choice for industrial applications. This idea was translated into precision agriculture field, designing platforms that consider hardware requirements and the conditions where the system will be placed to ensure robustness in the operation and low cost for the developed embedded platform. Field Programmable Gate Arrays (FPGAs) have been demonstrated to be a solution with a high performance, flexibility and robustness for greenhouse embedded applications (Castañeda-Miranda et al., 2006).

New tendencies on greenhouse climate control systems
This review clearly shows that there is a tendency to utilize climate controllers for protected agriculture applications where these are based on very simple control theories, like ON/OFF, PID controllers or some variation thereof. This tendency has been caused by the low computational power availability in the past. Consequently, the control was limited to basic operations and real-time processing was unreachable because of the absence of adequate processing devices. Because of this, recent mathematical algorithms and control theories evolved faster than computing technologies. Consequently, complex algorithms were not available to be implemented in any technological platforms at affordable cost a few years ago.
This scenario changed when computers cost decreased and the processing capabilities became considerably higher, at least to the point to make it possible to implement complex algorithms. Soon thereafter, modern control theories based on real-time control process, adaptive schemes or intelligent techniques were applied in order to achieve a more accurate, efficient and strict manipulation of the interest greenhouse variables. Considerations regarding quality, yield, water and energy savings were also studied and integrated in the control models (Vazquez-Cruz et al., 2010). These low cost platforms also make it possible to design and implement sensor networks, mobile robots for agricultural proposes, image processing for early diseases and pest detection as well as many other contributions to agriculture Contreras-Medina et al., 2009).
Despite the excellent results obtained with the high performances low cost computers, and advanced algorithms, tendencies are changing again; recently investigations and reports lead to the development of controllers which also consider plant physiology and morphology. Phenomena such as transpiration and photosynthesis have been studied for better understanding of plant behaviors in order to control climatic and nutritional requirements, according with real plant needs. New energy savings strategies have been proposed considering the available information regarding plant processes, manipulating climatic conditions when it is necessary for plant growth and the establishment of adaptive operating ranges for actuators with more degrees of freedom where strategies are helpful and when the reduction of energy consumption is essential. Nevertheless, the information about plant processes is limited. More investigation is necessary to establish correlations between physiological processes and plant growth in regards to temperature, humidity, nutrition and others controllable variables of the greenhouse, with the objective of reaching a sustainable protected agriculture industry (Millan-Almaraz et al., 2009).

Conclusions
Greenhouse climate control is currently one of the main objectives of engineering in precision agriculture. Temperature and humidity are variables which have a direct relationship with the plant production. Moreover, recent investigations have shown that is not enough to adjust temperature and humidity ratings to maximum and minimum setpoints which are affordable for plant needs. Because of this, many control theories have emerged along the years such as conventional control techniques and optimal control. Conventional control is based mainly on the proportional-integral-derivative controller and some variants. Furthermore, optimal control techniques relies mainly on AI algorithms and adaptive control theory which proposes an alternative way to solve the climate control problem when green-house mathematical model is unknown or often very complex. Another important fact which has limited the development of more advanced climate control system was the technological limitations a few years ago. However, the relatively new computational technologies such as microprocessors, digital signal processors and field programmable gate arrays are allowing continuing the implementation of more sophisticated control systems. According to some authors, the application of advanced controllers capable of following specific variable setpoints has not yet proven to be an optimal solution. Because of this, new tendencies are appearing in greenhouse climate control based on gathering extra information about physiological and morphological processes on the plant such and transpiration, stomata conductance and photosynthesis. These new control theories report that it is not necessary to have strict temperature and humidity set points. Instead of this, more flexible thresholds are proposed to save unnecessary energy consumption which is consumed when controller tries to follow the set point in a strict way. Phytocontrol is the new theory which proposes the use of the plant physiological responses as input sensor to establish the set point in the climate controller. Also, this has not proved to be a stable and reliable method, because it is necessary to gather a lot of information to prove the reliability of this. Nevertheless, different types of controllers have emerged demonstrating advantages and disadvantages between them, better performance for some actions among other characteristics. Researchers need to analyze different control theories to determine which one is the most proper for their projects according to their specific requirements of greenhouse climate control systems.