Anda di halaman 1dari 15

Applied Soft Computing 62 (2018) 86–100

Contents lists available at ScienceDirect

Applied Soft Computing


journal homepage: www.elsevier.com/locate/asoc

Modeling of a greenhouse prototype using PSO and differential


evolution algorithms based on a real-time LabViewTM application
Alfonso Pérez-González ∗ , Ofelia Begovich-Mendoza, Javier Ruiz-León
CINVESTAV IPN Campus Guadalajara, Zapopan, Jalisco, Mexico

a r t i c l e i n f o a b s t r a c t

Article history: In general, to obtain an adequate mathematical model of a greenhouse is a difficult task due to the
Received 7 November 2015 complexity of the involved equations that describe the dynamics of the system, and the required high
Received in revised form 30 June 2017 number of physical parameters, which are complicated or even impossible to measure. In these situa-
Accepted 13 October 2017
tions, estimation methods are commonly used to obtain a suitable approximation for those parameters.
Available online 20 October 2017
This paper presents the application and comparison of a collection of methods based on Particle Swarm
Optimization (PSO) and Differential Evolution (DE), using them as the tools to identify the parameters
Keywords:
that complete a proposed mathematical model for a greenhouse. These parameters are sought aiming to
Particle Swarm Optimization
Differential Evolution
approximate the dynamic behavior of a greenhouse physical prototype building in CINVESTAV Campus
Greenhouse model Guadalajara, by using the heuristic algorithms in order to minimize a proposed error function, which
LabViewTM application considers as arguments estimations and measurements of the two more representative dynamics of the
Real-time application climate conditions inside a greenhouse: namely, the air temperature and relative humidity. Different
forms of PSO and DE algorithms are considered and applied in order to select the one that achieves the
set of parameters with the lowest evaluation error. The comparison of the selected algorithms is carried
out in an offline optimization schedule using real data recorded through the LabViewTM SignalExpress
application, and a real-time implementation in a LabViewTM code, implemented to optimize the model in
a sample to sample execution. The proposed model, with its corresponding computed parameters, is val-
idated comparing its results against the real dynamic behavior of the temperature and relative humidity,
that are measured directly from the greenhouse prototype, showing a good agreement between real and
estimated values. Several tests were executed in order to find PSO and DE best calibration conditions.
Experimental results allow us to propose an efficient way to deal with numerical optimization problems
of high complexity, applying a two stages scheme based on a first offline pre-identification, where the
obtained results are used as initial condition for an online, real-time refinement process.
© 2017 Elsevier B.V. All rights reserved.

1. Introduction ural weather and presence of animals is ensured. The addition of


these features is in order to obtain a constant improvement in vol-
A greenhouse usually consists of a structure, build in metal in ume and quality for the crop production.
most cases, coated with a transparent material which provides the Regarding the greenhouse operation, an adequate mathematical
property to convert the building into a closed room, for the contain- model that describes the dynamic behavior of the system is often
ment of a growing area for plants. The main purpose of a greenhouse required for the purpose of control and manipulation of the climate
is the climate adaptation at the interior of the closed structure, variables, mainly when automatic control strategies are considered
aiming to produce the ideal conditions for the growth and devel- to be applied. Obtaining such model is a task of high complexity
opment of specific crops, providing independence of the external that depends on the intrinsic physical characteristics of the case of
environment while a level of isolation and protection against nat- study.
Generally speaking, it is common to think of two types of mod-
eling methodologies in order to describe the dynamic behavior
∗ Corresponding author. of a greenhouse prototype [3]. The first one is called Black-box
E-mail addresses: aperez@gdl.cinvestav.mx (A. Pérez-González), Based Modeling, which is a technique that attempts to approx-
obegovi@gdl.cinvestav.mx (O. Begovich-Mendoza), jruiz@gdl.cinvestav.mx imate the behavior of a greenhouse using data mining analysis
(J. Ruiz-León). or input/output measurements taken directly from the process.
URL: http://www.gdl.cinvestav.mx (A. Pérez-González).

https://doi.org/10.1016/j.asoc.2017.10.023
1568-4946/© 2017 Elsevier B.V. All rights reserved.
A. Pérez-González et al. / Applied Soft Computing 62 (2018) 86–100 87

This is a useful technique applicable in several ways, like ARMAX correspondence of the model estimations against the real mea-
type algorithms, polynomial adjustments, training of artificial neu- surements, in each actualization point. The experimentation with
ral networks [4], fuzzy logic implementations [5], etc. However, a variety of algorithms is in order to have a complete picture of
this kind of technique has the disadvantage that its results are the performance of each one, aiming to select the one with the
only numerical, without a physical correspondence. The second best characteristics for each stage of the methodology. We also
methodology is known as modeling based on acknowledgement, present a comparative study of different approaches reported in the
which implies an a priori specialized knowledge of the dynam- literature related to greenhouse modeling and the corresponding
ics and properties of the system. This technique seeks to model parameter identification, in order to establish the main differences
the implied dynamics using differential equations, where the between these methods and the proposed methodology, highlight-
involved parameters acquire a physical meaning, unlike the Black- ing the results obtained in this work.
box approach. It should be noted that obtaining such models is a Measurements of climatic variables were acquired using
very complicated task, due to the physical characteristics of the sys- LabViewTM , and the temperature and relative humidity inside the
tem, for instance, the presence of strong nonlinearities, stochastic greenhouse prototype are considered as main system dynamics,
behavior and dynamic delays [6]. Therefore, the models reported in unlike [12,16], where the relative humidity is not considered. In this
the literature usually take into account only the essential variables research, a different form of PSO than the one used in [12,14,16] is
to regulate, in general, the temperature and humidity inside the implement in order to obtain better results. This PSO form incor-
prototype [7–9], trying to simplify the resulting representation. porates the use of a convergence factor and a stability criterion.
An important problem appearing in acknowledgement mod- In a similar way, we show that by using good calibration param-
els is the involvement of a large number of unknown parameters, eters in the algorithms, we can avoid the use of a complicated
which requires a complicated process of instrumentation and meta-structure based on the Kalman Filter in order to improve the
experimentation in order to find their appropriate values. Never- heuristic search. As far as the programming language selection is
theless, there exists an increasing number of proposals to handle concerned, it is worth mentioning that LabViewTM has proven its
this situation, based on numeric algorithms which try to avoid efficiency in an enormous variety of fields, as it has been exposed,
experimental issues, as for instance: subspace methods [10], linear for instance, in [18,19], where the authors present environments
regressions [1], auto-regressive models [11], etc. Also, the use of for the visualization and data management of green energy sys-
heuristic algorithms as a new methodology to solve the high com- tems with high levels of instrumentation. We can also add that
plexity problem of finding the parameters for acknowledgement LabViewTM is a highly understandable programming language due
models has recently been proven to be an effective solution. In par- to its graphical interface, and it is equipped with several tools
ticular, the Particle Swarm Optimization (PSO) algorithm has been similar to pre-programmed macros that makes the design of appli-
tested as a simple and effective methodology to deal with the opti- cations much easier. In addition, the philosophy of modular code
mization of complex functions with a large group of parameters, as construction of LabViewTM allows to design and diagnose the code
shown, for instance, in [12–17]. in an easier way. Also, LabViewTM is the natural programming lan-
The aim of this work is to apply, compare and validate the guage for the device used in this research as controller and data
efficiency of a variety of heuristic techniques, to be more spe- logger, a CompactRIO, which is an equipment with its own real-
cific, variations of both, the Particle Swarm Optimization algorithm time OS, a feature that can be exploited in order to reduce the
(PSO) and Differential Evolution algorithm (DE), in the task of non-deterministic execution time for computational tasks, called
obtaining an estimation of the parameters of a proposed green- jitter.
house model [12], and its adaptation to a real study case. Parametric The structure of this paper is the following: Section 2 presents
estimation of greenhouse models using heuristics has also been the description of the greenhouse prototype. Section 3 introduces
considered in [12,14,16,17], as works closely related to the research the selected mathematical model and the set of considerations to
of this paper. Particularly, in [12] the authors present a model that adapt it to the case of study of this research. Section 4 describes
takes into account the temperature and the water vapor pressure, the identification problem. Section 5 presents the forms of PSO
but in practice and due to instrumentation requirements, it is very and DE algorithms that will be applied and compared. Section
complicated to measure the water vapor pressure. Instead of that, 6 shows the implementation of the algorithms, with the respec-
an adaptation to that model is applied in the present work, in order tive results for every case, taking into account offline and online
to use the temperature and relative humidity dynamics, as it has versions. Section 6 also presents some comments about different
been proposed originally in [17] with good results. In a similar way, methods in the related state of the art, in order to establish com-
in [14] the authors present a model that considers the temperature parison points between these methods and the results obtained in
and relative humidity dynamics, but they only present an offline this work. Finally, we offer the relevant conclusions about all the
optimization method with a considerable approximation error. In implementation.
[16] the authors introduce the use of an Extended Kalman Filter
and a neuronal network in the application of a PSO algorithm, but
2. Greenhouse prototype
they only considered the internal greenhouse temperature. Finally,
in [17] the authors present a full methodology based on heuristic
The greenhouse prototype considered in this study is located
methods for the parameter search of a greenhouse mathematical
at CINVESTAV (Center for Research and Advanced Studies) in
model, but a comparison between a collection of algorithms is not
Guadalajara, México (Global Position Coordinates at 20.668505,
addressed.
−103.465039). At the moment of this research, the prototype was
The main contribution of this work, compared with the above
in a stage where it does not have any kind of crop, considering the
mentioned references, is the application and comparison of a vari-
recommendations of biological engineering experts, who suggest
ety of PSO and DE algorithms to estimate the parameters of a
that it is important to characterize first the thermodynamic behav-
proposed mathematical greenhouse model, and the validation of
ior of the greenhouse in an empty stage. The greenhouse prototype
the results on a real greenhouse, with the intention of estab-
has a floor area of 30 m2 and a maximal height of 5.14 m, as it is
lishing a schedule of optimization of high performance. Results
shown in Fig. 1. It is also equipped with:
obtained in both, the offline and online scenarios, allow to propose
a new methodology based on a pre-identification stage, followed
by an online sample-to-sample search process that maintains the • Internal and external temperature sensors.
88 A. Pérez-González et al. / Applied Soft Computing 62 (2018) 86–100

Fig. 1. Schematic representation of the greenhouse prototype (all measures are in meters).

• Internal and external relative humidity sensors. • Y(t) as the set of observed/measured states,
• Internal and external wind speed sensors. • X(t) as the actual states vector,
• Internal and external CO2 concentration sensors. • U(t) as the set of system inputs, and
• Internal solar radiation sensor. • P as the vector of unknown system parameters.
• Two lateral windows, each one with a surface of 11.4 m2 and
equipped with an opening control system that uses an electric Thus, the dynamic behavior of the system can be represented
engine of 100 W, 100 Nm with speed of 3.6 RPM, including an by the equations
E50S incremental optical encoder codified in X2.
dX(t)
• Two fans of 1/4 HP and 750 m3 /h each one. = f (X(t), U(t); P), X(0) = X0 (1)
• dt
UV2 plastic coat of 70% of solar permeability.
• Roof window with a surface of 4.34 m2 . Due to the lack of an Y (t) = g(X(t), U(t); P), (2)
automation system for this window, it was always closed for all
where (1) is the state equation and (2) is the observation equation.
the presented experiments and results.
For the sake of readability of the next paragraphs, a glossary of the
• NI c-RIO-9074 as data acquisition equipment.
variables and parameters of the greenhouse, including their units,
is presented at the end of this paper.
Considering all the sensors, the one which determines the sam- According to [12], in order to reduce the complexity of the
pling conditions is the CO2 sensor, since it has the slower time of model, it is necessary to consider an empiric approximation based
actualization of all the sensors set. In fact, the CO2 sensors have on the system thermic components, like the soil and the heavy
an actualization time of 1 s, after an initial transient state of 5 s, structural parts, taking then the virtual thermic mass with temper-
in which the sensor reaches its operation point. Due to that, the ature Tm and calorific capacity Cm . Also, the air volume contained
selected sampling time for all the experiments is taken as 1 s, in at the interior of the prototype constitutes another representative
order to avoid redundant data in the CO2 sensors case, considering element of the thermic interactions inside the greenhouse. Thus,
also that all the greenhouse dynamics are of slow behavior (except we can characterize the air volume inside the prototype according
for the wind speed). Then, although all the other sensors have to its temperature Ti and the relative humidity percentage Hi . The
shorter actualization times, the dynamic behavior of their mea- virtual thermic mass equation has the form
sured variables varies poorly in these intervals, so, 1 s as sampling
time allows to establish a good compromise between the informa- dT m
Cm = h(Ti − Tm ) + Qsol + ˇRg , (3)
tion provided by the measurements and the size of the conformed dt
data base. The wind speed is also conditioning to this sampling where h(Ti − Tm ) represents a thermic exchange between the air
time, since it really does not influence the model considerably. and the virtual thermic mass, Qsol represents an artificial thermic
flow on the floor of the greenhouse, and finally, ˇRg represents a
3. Greenhouse model proposal gain due to the solar radiation incidence.
The thermic balance of the air inside the prototype, neglecting
Greenhouse modeling techniques require approximations of the thermic inertia, obeys the expression
thermic and mass transferences. This can be represented by five 0 = ˛Rg + Qair + h(Tm − Ti ) + K(Te − Ti ) + Ks (Te − Ti )
variables that, in general, describe the behavior of the greenhouse
prototype and its environment: +Kl (He P(Te ) − Hi P(Ti )), (4)

where ˛Rg is a thermic gain due to the solar radiation, Qair


• X0 as the initial conditions vector, is an artificial thermic input, h(Tm − Ti ) represents the thermic
A. Pérez-González et al. / Applied Soft Computing 62 (2018) 86–100 89

exchange between the air and the virtual thermic mass, K(Te − Ti ) Table 1
Physical constants involved in the greenhouse model.
represents the thermic interaction between internal and exter-
nal air of the greenhouse prototype, and finally, Ks (Te − Ti ) and Greenhouse known parameters
Kl (He P(Te ) − Hi P(Ti )) are the sensitive and latent heat exchange due
Parameter Symbol Value
to ventilation and leakages, respectively.
Air thermic capacity Cp 1002.43
An adaptation to the model presented in [12] is proposed next,
Greenhouse volume v 122.52
where we introduce a relative humidity characterization instead of Sampling time t 1
the water vapor pressure dependency, in order to avoid the pres- Air density 1.297
sure instrumentation, which is more expensive and less available,
in comparison with relative humidity sensors; also, the environ-
ment specifications for crop development are generally expressed
where
in temperature and relative humidity. According to [20], the factor
H* P(T* ) represents a suitable approximation of the pressure gener- ⎡ ⎤T
1
ated by the presence of water vapor from the Clausius–Clapeyron
⎢ ˛h + ˇv ⎥ ⎡T ⎤
curve ⎢ h(K + Ks ) ⎥ e(k)

⎢ ⎥ ⎢ Rg(k) ⎥
P(T∗ ) = 6.11 × 10 T∗ +237.3 .
7.5T∗ ⎢ v ⎥ ⎢ ⎥
(5)
−t
⎢ ⎥
−t ⎢ h(K + Ks ) ⎥ ⎢ Qsol(k) ⎥
⎢ ⎥ ⎢ ⎥.
Therefore, the use of Eq. (5) allows to obtain an approximation Tm(k+1) = Tm(k) e  + (1 − e  )
⎢ 1 ⎥ ⎢ ⎥ (13)
⎢ (K + Ks ) ⎥ ⎢ Qair(k) ⎥
of the relative humidity dynamics, using as a basis the pressure
⎢ K ⎥ ⎢ ⎥
generated by the water vapor presence, as follows: ⎢ l ⎥ ⎣ He(k) P(Te(k) ) ⎦
⎢ (K + Ks ) ⎥
Pi = Hi P(Ti ), Pe = He P(Te ). (6) ⎣ ⎦ Hi(k) P(Ti(k)
−Kl
The balance of water vapor inside the prototype can be repre- (K + Kl )
sented by
dP i 4. Model identification
Cl = A Rg + B(P(Ti ) − Pi ) − Kl (Pi − Pe ) + ϕl , (7)
dt
dP For the estimation of the temperature for the virtual mass, the
where Cl dti represents the dynamic variation of the environmen- relative humidity and the temperature inside the greenhouse pro-
tal pressure at the interior of the greenhouse, A Rg is a linear
approximation of the crops transpiration due to the solar radiation, totype at time k + 1, the corresponding notation is T̂ m(k+1) , Ĥ i(k+1)
B(P(Ti ) − Pi ) represents the air saturation deficit due to the pres- and T̂ i(k+1) , respectively. Also, T̂ i(1) = Ti(1) and Ĥ i(1) = Hi(1) are taken
ence of water vapor, Kl (Pi − Pe ) is the ventilation exchange between as the initial conditions of the system, where Ti(k) and Hi(k) are the
the interior and the exterior of the greenhouse prototype, and ϕl internal temperature and relative humidity, respectively, measured
represents an artificial contribution of water vapor. directly from the prototype.
According to [12], by simultaneous integration of Eqs. (3) and According to [14], the parameters presented in Table 1 are suit-
(4), and considering Eq. (7) in terms of Eq. (6), a three equations able.
system is obtained, which includes three unknown values, Tm , Ti Due to the building stage of the prototype, the actuators terms
and Pi . This system is represented in a recursive form that depends Qsol(k) , Qair(k) and ϕl(k) are neglected in order to represent the lack
on the passed calculations, the set of inputs {Rg , T0 , V, P0 , P(Ti )}, of these instruments. Finally, the set of unknown parameters of the
the set of actuators {Qsol , Qair , ϕl , s}, and the model parameters that model is presented in Table 2, where it is also included the expected
need to be identified. physical bounds for each one.
Thus, the internal pressure dynamics is represented as follows:
⎡ ⎤
Rg(k)
 ⎢ ⎥
rSB  (k) SB S ⎢ He(k) P(Te(k) ) ⎥
Pi(k+1) = Hi(k) P(Ti(k) )e−(k) t + (1 − e−(k) t ) ⎢ ⎥ ,(8)
(k) (k) (k) (k) ⎢ P(T ) ⎥
⎣ i(k) ⎦
ϕl(k)
where
√ √
Al sV (k) C + Al s0 V(k) C + d0 + BS/ Cp
(k) = , (9)
v
√ √
(k) = Cp Al sV (k) C + Cp Al s0 V(k) C + Cp d0 + BS, (10) In order to apply a heuristic optimization algorithm, it is neces-
sary to propose a cost function whose values represent a numerical
and interpretation of the performance of the implemented method.
(k) = (k) − BS. (11) Considering the internal temperature and relative humidity as the
most representative climatic variables for the greenhouse proto-
In the same way, using (3) (4), (5) and (6), the temperature type, the following Mean Square Error (MSE) function is established
dynamics inside the greenhouse is expressed as follows: as the cost function:
⎡ Te(k)

1 
N
⎢ ⎥ 2 2

⎢ Rg(k)
⎥ J= [(T̂ i(k) − Ti(k) ) + (Ĥ i(k) − Hi(k) ) ], (14)
h v−h ˛ 1 Kl −Kl ⎢ ⎥ 2N

⎢ ⎥
Ti(k+1) = Tm(k+1) + b Qair(k) (12) k=1
v


⎢ ⎥
⎣ H e(k) P Te(k) ⎦
where the arguments of this function are the estimations and mea-
H i(k) P Ti(k) surements of these representative variables.
90 A. Pérez-González et al. / Applied Soft Computing 62 (2018) 86–100

Table 2
Unknown parameters and their boundaries for model optimization.

Greenhouse unknown parameters

Parameter Symbol Lower bound Upper bound

Thermic mass initial temperature Tm(1) 0 50


Thermic exchange convection coefficient h 0.5 40
Coefficient of absorption of solar radiation due to air ˛ 0.2 100
Characteristic time  1 2000
Coefficient of absorption of solar radiation due to mass ˇ 0 50
Transpiration coefficient B 1 100
Windows natural ventilation factor Al 0 100
Natural ventilation factor due to air convection C 0 5000
Leakage surface s0 0 1000
Leakage factor d0 0 200
Thermal loss coefficient K 0.001 5000
Latent heat transfer coefficient Kl 0.0001 200
Sensible heat transfer coefficient Ks 0.0001 2000
Mix ratio r 0 1000
Exchange surface between components S 0 122
Psychrometric constant  0 100
Coating transitivity coefficient  0 1000

5. PSO and DE algorithms numbers. The output of the function f(.) is also a real number that
measures the performance of its argument. Thus, the best position
Heuristic methods are considered as a useful tool to solve a great of each particle as a candidate solution corresponds to the mini-
variety of optimization problems, mainly where the process of find- mum value of the function f(.), now seen as the objective function.
ing an analytic solution represents a big effort or an enormous The aim of the evaluation process is to find a candidate solution ˛
consumption of time. Thus, this kind of algorithms are a suitable such that f(˛) ≤ f(a) for all a in the search space, which implies that
method to be applied when the problem to be solved admits a ˛ is the argument that minimizes the objective function globally.
lack of exactitude in its final result, but time saving or less effort is Let PRN be the total number of particles in the swarm (candidate
required in the calculations. solutions), each one provided with a position PRi ∈ Rn in the search
space, and a speed VPRi ∈ Rn ; and let Pbest i ∈ Rn be the memory
5.1. PSO selected forms of the best position known for the ith particle, and let Gbest ∈ Rn
be the memory of the best global position known (between all
(PSO), as one of the most well-known heuristic algorithms, has particles). Then:
been proven in an extensive field of applications as an effective For each particle PRi , i = 1, . . ., PRN:
method to obtain a suitable solution for optimization problems.
It is also a common source of inspiration for the creation of more • The position of each particle is initialized by a random uniform
complex hybrid methods, as we can see, for instance, in Please, omit distribution PRi ∼ U(LB, UB), where LB and UB represent the lower
the [20–22]. PSO presents a series of characteristics that makes it bound and the upper bound, respectively, of the search space.
the first choice in the algorithm selection for this research: • The best known position of each particle is initialized using the
position obtained in the previous step, Pbesti ←− PRi ; also, the
• PSO has a real valued representation, which allows to avoid the global best known position is initialized with the value of a random
conversion to binary field and backwards, common in many member of the set of particles Gbest ←− PRrandI[1,PRN] .
heuristic algorithms. • If f(PRi ) < f(Gbest), then the global best known position is updated
• The set of operations that compose the PSO algorithm is based on by Gbest ←− PRi .
arithmetic, and it is of easy understanding. • The velocity for each particle is initialized using a random uniform
• PSO presents a swarm behavior, which is an adequate approach distribution VPRi ∼ U(−|UB − LB|, |UB − LB|).
in search spaces of considerable extension, due to the capability
of exploration in steps of different length and the communication After the above steps, the algorithm becomes iterative, and it
between particles, which share the information of the best found will be necessary to establish a stop criterion, generally given by
results. a maximum number of iterations. Then, while the stop criterion is
• PSO is well known, thus, there exists a lot of publications about not reached, it is necessary to repeat, for each particle PRi , i = 1, . . .,
it, and many variations have already been proposed, in order PRN, the velocity update, the position update and the evaluation
to tackle problems of considerable complexity. Besides, there process.
already exists references about PSO calibration and stability. In the current literature there exists a large number of pro-
posals for PSO variations. For this work, 3 representative forms
PSO is a heuristic computational algorithm based on group were selected. The first one, the so-called Classical PSO, has a speed
search behavior that allows to optimize the solution of a problem update equation as follows [21,22]:
through a set of candidate solutions [21]. PSO assumes a few issues
or no assumptions at all about the problem to be optimized, and it VPRi,d ←− ωVPRi,d + ϕp rp (Pbest i,d − PRi,d ) + ϕg rg (Gbest d − PRi,d ).
can be applied in large search spaces, as mentioned above. Still, it (15)
is impossible to guarantee an optimum solution in all cases.
The PSO algorithm will be briefly described next: take f (a) :
Rn → R as a real valued cost function required to be minimized, The second PSO implemented is the form named Many Optimiz-
where a represents a candidate solution taken as a vector of n real ing Liaisons PSO (MOLPSO) [23,24], proposed as a simplification of
A. Pérez-González et al. / Applied Soft Computing 62 (2018) 86–100 91

the original algorithm. It eliminates the self-conscience of each par- has or not better performance than greedy search in problems like
ticle, suppressing the memory of the best known position Pbesti . Its the one presented in this research.
speed update equation takes the form: • DE is composed by arithmetic operations that allow to create
easy representations of the set of candidate solutions, and easy
VPRi,d ←− ωVPRi,d + ϕg rg (Gbest d − PRi,d ). (16) interactions between these candidate solutions.
• DE only needs two calibration factors, and the definition interval
The last selected variation is the so-called Increased Convergence
for these factors is considerably small. This defines DE as an easy
Factor PSO (ICFPSO); it is a Classical PSO with the addition of a com-
calibration algorithm.
pression factor Cs , taken as the convergence factor. In practice, it is
proven that ICFPSO has a greater convergence speed than the regu-
lar PSO form [22,25]. The equation for the speed update for ICFPSO DE algorithm, like PSO, is a computational algorithm based on
has the form: the manipulation of a population of candidate solutions, applicable
for complex search problems [26]. It presents a series of interactions
VPRi,d ←− Cs (ωVPRi,d + ϕp rp (Pbest i,d − PRi,d ) + ϕg rg (Gbest d − PRi,d )). between the candidate solutions to produce new individuals, and
such new members are tested and catalogued by a cost function,
(17)
seeking for the survival of only the ones with the best performance.
The main characteristic of the DE algorithm over other evolutive
For all the presented PSO variations, d = 1, . . ., n represents the methods, is the use of trial vectors of real values, as mentioned
index of the dth element of each n-dimensional particle PRi ; ω rep- before.
resents the particle inertia, which modulates the influence of the Let PR be a set of vectors, where the ith member is denoted by
past speed of each element of the particle in the latest update; ϕp is PRi . Besides, PRi ∈ Rn , thus, the dimension of each vector, n, equals
the local attractor, also called cognitive factor, which modulates the the number of parameters that are needed to be identified. PR is a
search capability of new areas; and ϕg is the global attractor or social finite set, formed for all the PRi ∈ PR, i = 1, . . ., PRN.
g
factor, which controls the fine-search capability of the algorithm in Thus, PRi,m is an element, or gene of a member of the population,
the neighborhood of the global best known position. The factors rp such that i represents the individual index inside the population, m
and rg are uniformly distributed random gains for the local and the represents the index for each gene of a vector PRi , with m = 1, . . ., n,
global attractors, respectively, selected in the interval [0, 1]. and g is the generation to which the individual belongs, that is, an
According to [22,25], the next expressions are considered for a element with g = x + 1 is the direct product of the genetic interac-
good performance and stability of ICFPSO: tions of members with g = x. Also, let LBm and UBm be the minimum
and maximum bounds, respectively, for all the genes with index m.
ϕp + ϕg = ϕ > 4, (18) The initial population, g = 1, is generated by the expression
and PR1i,m = LBm + re (UBm − LBm ), (21)
2
Cs =  . (19) where re takes values from a random uniform distribution re ∼ U(0,
|2 − ϕ − ϕ2 − 4ϕ| 1). After the initialization, a recursive mutation process is per-
formed over the population, and it is the beginning of each iteration
After performing the speed update, the next step is to update
during the DE algorithm execution. The form of the mutation is the
the particle position using the expression
part of the method that defines the DE algorithm variation. The
PRi ←− PRi + VPRi . (20) mutation expressions are presented later on.
The mutation process generates a set of mutant individuals,
g
To finalize each iteration, if f(PRi ) < f(Pbesti ), the best known posi- VPRi , i = 1, . . ., PRN, then, a recombination stage is executed, using
tion of the ith particle is updated by Pbesti ←− PRi , and, in the same the expression
way, if f(PRi ) < f(Gbest), the global best known position is updated  g
by Gbest ←− Pbesti . When the stop criterion is reached, Gbest is VPRi,m if rv < Cr or m = IrandI[1,n]
g
Ui,m = (22)
presented as the best solution found. Notice that the evaluation g
PRi,m otherwise,
function f(.) corresponds, in our case of study, to Eq. (14).
It is important to point out that the Classical PSO is the first g
where Ui is the trial vector, and it will be the argument of the cost
implemented algorithm of this research, allowing the selection function during the evaluation step; Cr , known as recombination
of MOLPSO as a comparison between the well known PSO and a crossover, directly modulates the form of the combination between
simpler form; the selection of ICFPSO is in order to compare the g
the genes of the original member PRi and the ones from VPRi ;
g
influence of the calibration process presented by Clerc and Kennedy IrandI[1,n] is a random positive integer, whose condition guarantees
[25] in the obtained results. g g
Ui = / PRi as long as there exists genetic diversity in the population;
and rv takes values from a random uniform distribution rv ∼U(0, 1).
5.2. DE selected forms Each iteration finishes in the evaluation stage, where the popu-
lation is restructured using the expression
Differential Evolution (DE) represents the second choice for this  g g g
research. As the PSO algorithm, DE has been proven as a good per- Ui if f (Ui ) < f (PRi )
g+1
formance method in the approximation of solutions for problems of PRi = (23)
g
PRi otherwise,
high complexity. The characteristics that makes it a suitable option
are enlisted below:Replace D̈Ifferential Evolution (DE)ẅith only D̈E¨, where, as in the PSO algorithm, f(.) is given by Eq. (14). The stop
please. criterion also uses an iteration number goal.
An elitism criterion is added to each iteration by the expression
• Like PSO, DE needs no conversion schedule for the representation  g g
of the population. It allows to track the problem directly in Rn . Ui if f (Ui ) < f (Gbest)
• DE is considered as a greedy search algorithm. This marks a differ- Gbest = (24)
Gbest otherwise,
ence between PSO and DE, and allows to check if swarm behavior
92 A. Pérez-González et al. / Applied Soft Computing 62 (2018) 86–100

where Gbest is presented as the best solution found at the end of refers to Mean Square Error between the temperature and relative
the execution. humidity measurements and their respective estimates.
Recalling the mutation stage, the first selected form for this Fig. 3 shows the graphical results for the internal temperature
research is the so-called DE rand/1 [27] as the most common form and relative humidity estimations using the parameters found in
of DE mutation. It uses a single random combination of elements the last iteration of the Classical PSO.
from the original population to generate each mutant, as expressed The results for MOLPSO are shown in Table 4. For this algorithm,
in the following equation: calibration parameters were selected according to [29], where the
g g g g increment of the population number is taken as a technique to
VPRi = PRr1 + F(PRr2 − PRr3 ), (25) tackle problems with a high number of parameters to be identified,
where r1 = / r2 = / r3 are random integers uniformly distributed in at the same time that this increment also compensates the unde-
the interval [1, PRN]. sirable effects of the suppression of the self-conscience of MOLPSO.
The second choice of mutation method is the one known as DE Also, it is important to point out that MOLPSO does not present a
current to best/1 [28], because it uses a distance between the mem- fast premature convergence behavior, that is, it does not require a
ber that is indexed at the moment of mutation and the element schedule of multiple steps to recalibrate the algorithm.
of best performance in the whole population at the g generation, Fig. 4 shows the graphical results for the internal temperature
and it also includes a random combination of elements, like in the and relative humidity estimations with the model that includes
previous case. The expression for this mutation form corresponds the parameters obtained by MOLPSO, the improvement is notable
to in the MSE presented in Table 4 and in the graphic representation,
in comparison with the Classical PSO results.
g g g g g
VPRi = PRi + F(Gbest − PRi ) + F(PRr1 − PRr2 ), (26) For the last PSO form, the ICFPSO, the parameters were selected
with an experimental criterion based on the best results obtained
where r1 = / r2 are random integers uniformly distributed in the
from several executions, considering the Eqs. (18) and (19). As in
interval [1, PRN]. For both mutation cases, F represents the mutation
the MOLPSO implementation, ICFPSO does not present the fast pre-
factor, and it is a real constant to be fixed in the interval (0, 2), which
mature convergence behavior. In fact, it was not possible to improve
regulates the algorithm capability of exploration (values near to 2)
the obtained results reached after 1800 iterations for none of the
and exploitation (values near to 0).
PSO forms presented in this paper, neither adding a recalibration
DE rand/1, like Classical PSO, is selected as the most common
step, nor increasing the size of the population or performing more
form of DE algorithms, allowing DE current-to-best/1 to be imple-
iterations. Results for ICFPSO are presented in Table 5.
mented to check the influence of social interaction in a greedy search
Fig. 5 shows the behavior of the model with the values obtained
algorithm.
from ICFPSO as parameters, comparing the estimations of the
internal temperature and relative humidity with their respective
6. Application and results measurements.
Experimental results show a significant reduction of the MSE
For the computational implementation, the deployed equip- between the Classical PSO and the ICFPSO (7.1214 units), while, for
®
ment is an Intel CoreTM i7-2600 K, with a 3.4 GHz processor and the MOLPSO there exists a small MSE difference in comparison with
15.7 GB in RAM, DDR3 833 MHz type; the OS of the computer ICFPSO (0.426 units). The execution time is almost the same for all
is WindowsTM 8 professional edition, and the machine is also the algorithms. Referring to the execution time of each particle, the
equipped with an operative LabViewTM 2012 version software. less expensive form is obviously the MOLPSO, which represents the
The acquisition and control system is a NI-CompactRIO 9074 with 99.80% of the time consumption of the other two algorithms; but,
a Freescale MPC5200 industrial processor of 400 Mhz, real time due to the calibration parameters suggested in [29], the MOLPSO
®
Windows River VxWorks OS; and a reconfigurable FPGA XilinxTM population is 28 particles bigger than the other variations, thus,
Spartan III with 2,000,000 of logic gates. it takes more time in total execution. Besides, the obtained results
The experimental schedule consists, as the first step, on an clearly show that the PSO form with the best performance is the ICF-
offline test of the described algorithms. Using the LabViewTM Sig- PSO, therefore, it will be taken to compete against the DE algorithm
nalExpress toolkit, a data base of climatic conditions of 24 h length of best performance.
is recorded, starting on November 20, 2013, at 2:45 p.m. The data
acquisition has a sampling time of 1 s, and is captured with the pro- 6.2. DE implementation
totype windows closed and fans turned off. Fig. 2 shows the data
base. Proceeding in the same way as in the PSO algorithms, the cli-
matic data base shown in Fig. 2 is taken as test data for the proposed
6.1. PSO implementation DE variations. For the DE rand/1, Table 6 shows the results. Calibra-
tion parameters were selected based on the best results obtained
The first tested algorithm, the Classical PSO, is implemented in after several experimentation scenarios, where it is important to
3 stages, each one with different calibrations parameters. Exper- point out that, neither bigger populations nor extended iteration
imentally, it is possible to observe that the Classical PSO form goals improve the obtained results. Also, DE algorithms do not
presents a behavior of premature convergence for any calibration require recalibration steps throughout their execution, due to their
parameters. That is, PSO rapidly admits a local minimum as the capability of reaching excellent results without fast convergence
best solution found. This behavior can be minimized changing the behavior.
calibration parameters at the moment of premature convergence. Recalling the results of the ICFPSO, it is evident that DE rand/1
Thus, the result of the first stage is taken as the initial condition has a higher MSE result, but the execution time for this DE algorithm
for the second stage, and, in the same way, the result of the sec- is less than 12% of the ICFPSO execution time. Of course, this is due to
ond stage is taken as the initial condition for the third stage, every the low population volume requirement for the DE algorithm. The
time, adjusting the calibration parameters. For this research, after behavior of the mathematical model that includes as parameters
the third calibration step, there is not a fourth calibration that can the values that were found by DE rand/1 and its comparison against
improve the obtained results. Calibration parameters, execution the temperature and relative humidity measurements are shown
times and results are shown in Table 3, where, the MSE acronym in Fig. 6.
A. Pérez-González et al. / Applied Soft Computing 62 (2018) 86–100 93

Fig. 2. Data base for November 20, 2013, starting at 2:45 p.m.

Fig. 3. Measurements and estimations for Classical PSO.

The last proposed algorithm, the DE current to best/1, is imple- ously mentioned. At particle level, DE algorithms have an execution
mented in the same conditions that the other ones, and the obtained time that represents, for rand/1 form, 107.56% of the execution time
results are shown in Table 7. Calibration parameters, like in the of each ICFPSO particle, and for the current to best/1 form, the cor-
other algorithms, were selected experimentally, based on the best respondence is 108.86%. However, the DE forms are less expensive
results obtained after several tests. in general terms, and also the DE current to best/1 form has the
Graphical results for the DE current to best/1 are shown in Fig. 7. best performance of all the executed algorithms, with a MSE value
At this point, it is possible to present a brief discussion about the of 1.3721 units less than the result of the PSO form with the best
results and comparisons of the proposed algorithms. First, the exe- performance.
cution time for the DE forms is considerable less than for the PSO About the number of iterations, it was selected considering the
forms, due to the low population density requirement, as previ- best case of execution for each kind of algorithm, after several
94 A. Pérez-González et al. / Applied Soft Computing 62 (2018) 86–100

Table 3
Implementation of Classical PSO.

Calibration parameters Execution time Results

First stage
ω = 0.341 Execution time for each particle = 1.534 s Tm(1) = 19.8168; K = 498.2187;
ϕp = 2 Execution time for each iteration = 306.8 s h = 11.5137; Kl = 0.7423;
ϕg = 4 Total execution time = 25 h34 m ˛ = 0.2720; Ks = 50.2080;
 = 210.8597; r = 1.2497;
PRN = 200 ˇ = 0.1388; S = 17.4738;
itermax = 300 B = 4.0917;  = 10.6474;
Al = 4.2710;   = 0.3976.
C = 1.9601;
s0 = 0.4909; MSE = 118.5424
d0 = 1.0171;

Second stage
ω = 1.302 Execution time for each particle = 1.534 s Tm(1) = 36.834; K = 749.1368;
ϕp = 2 Execution time for each iteration = 306.8 s h = 0.254; Kl = 0.0008;
ϕg = 2.6 Total execution time = 85 h13 m20 s ˛ = 0.4167; Ks = 1.7701;
 = 77.4686; r = 8.11;
PRN = 200 ˇ = 1.5678; S = 0.0277;
itermax = 1000 B = 2.10967;  = 1.9146;
Al = 9.38;   = 17.6603.
C = 181.8587;
s0 = 0.8131; MSE = 38.1441
d0 = 0.4898;

Third stage
ω = 0.748 Execution time for each particle = 1.534 s Tm(1) = 43.3350; K = 533.0516;
ϕp = 2 Execution time for each iteration = 306.8 s h = 0.7201; Kl = 1.431;
ϕg = 2.3 Total execution time = 42 h36 m40 s ˛ = 0.4105; Ks = 14.4136;
 = 48.0071; r = 5.3423;
PRN = 200 ˇ = 0.2213; S = 0.3052;
itermax = 500 B = 1.4097;  = 5.0511;
Al = 5.9094;   = 10.9406.
C = 66.7576;
s0 = 7.4737; MSE = 17.9975
d0 = 1.5762;

Total iterations: 1800 Execution time for all stages: 153 h24 m

Table 4
Implementation of MOLPSO.

Calibration parameters Execution time Results

ω =−0.3747 Execution time for each particle = 1.531 s Tm(1) = 49.4233; K = 61.3820;
ϕg = 4.2373 Execution time for each iteration = 344.96 s h = 24.7059; Kl = 0.1108;
Total execution time = 172 h28 m55 s ˛ = 0.3822; Ks = 73.4544;
PRN = 228  = 970.0159; r = 0.8627;
itermax = 1800 ˇ = 4.3994; S = 2.5027;
B = 1.0279;  = 0.3293;
Al = 9.8001;   = 6.2293.
C = 22.2470;
s0 = 99.6807; MSE = 11.3021
d0 = 0.0614;

Fig. 4. Measurements and estimations for MOLPSO.


A. Pérez-González et al. / Applied Soft Computing 62 (2018) 86–100 95

Table 5
Implementation of ICFPSO.

Calibration parameters Execution time Results

ω = 0.87 Execution time for each particle = 1.534 s Tm(1) = 45.7231; K = 72.5843;
ϕp = 2 Execution time for each iteration = 306.8 s h = 22.5012; Kl = 0.0502;
ϕg = 2.54 Total execution time = 152 h24 m ˛ = 0.7342; Ks = 70.2445;
Cs = 0.4871  = 992.121; r = 0.9146;
ˇ = 4.1218; S = 2.0144;
PRN = 200 B = 1;  = 0.3373;
itermax = 1800 Al = 10.0139;   = 6.2175.
C = 27.6392;
s0 = 100.0231; MSE = 10.8761
d0 = 0.0426;

Fig. 5. Measurements and estimations for ICFPSO.

Table 6
Implementation of DE rand/1.

Calibration parameters Execution time Results

F = 0.8 Execution time for initial evaluation = 49.8 s Tm(1) = 49.9945; K = 44.1889;
Cr = 0.5 Execution time for each particle = 1.65 s h = 36.0319; Kl = 0.0491;
Execution time for each iteration = 49.5 s ˛ = 0.3944; Ks = 81.0269;
PRN = 30 Total execution time = 17 h53 m19.8 s  = 880.7339; r = 15.6669;
itermax = 1300 ˇ = 4.2109; S = 65.2484;
B = 1.8529;  = 9.5230;
Al = 89.4319;   = 17.1960.
C = 38.7875;
s0 = 92.5775; MSE = 10.8806
d0 = 0.3740;

Fig. 6. Measurements and estimations for DE rand/1.


96 A. Pérez-González et al. / Applied Soft Computing 62 (2018) 86–100

Table 7
Implementation of DE current to best/1.

Calibration parameters Execution time Results

F = 1, 2 Execution time for initial evaluation = 49.8 s Tm(1) = 48.9387; K = 137.4996;


Cr = 0.6 Execution time for each particle = 1.67 s h = 23.2731; Kl = 0.3015;
Execution time for each iteration = 50.1 s ˛ = 1.8341; Ks = 92.5755;
PRN = 30 Total execution time = 18 h06 m19.8 s  = 1064.6769; r = 213.3789;
itermax = 1300 ˇ = 7.0721; S = 15.107;
B = 2.9434;  = 0.0454;
Al = 6.886;   = 126.641.
C = 0.3478;
s0 = 6.6954; MSE = 9.5040
d0 = 32.1055;

Fig. 7. Measurements and estimations for DE current to best/1.

experiments. For the PSO with the best performance, the ICFPSO, needs a few changes: in first place, the measurement of climatic
the best result takes almost 1800 iterations in order to reach the signals needs to be performed directly by LabViewTM . The eval-
convergence of all the particles to a single value, after that, no uation scheme for particles also needs to be restructured, and in
improvement can be obtained. The exact value of iterations was order to guarantee the real-time execution, it is necessary to ensure
1784, so, we rounded out that number to 1800 and used it in all that the execution time of each algorithm is less than the sampling
the PSO forms. In fact, the Classical PSO reaches the premature time. Selecting the same sampling time as 1 s, the evaluation of the
convergence much sooner than that value, around 1600 iterations, algorithms takes 0.524 s for the ICFPSO with 100 particles per sam-
without an improvement of the estimated solution. In the same ple, and 0.1711 s for the DE current to best/1 with 30 particles per
way, MOLPSO reaches the consensus of all its particles near 1750 sample. A total execution time of 0.6951 s is suitable for a period
iterations. The DE algorithms case is similar: DE current to best/1 between samples of 1 s, and with the minimum jitter characteristic
reaches the consensus of all its population in exactly 1298 iter- of a CRIO system, the implementation is guaranteed to be in real-
ations, while DE rand/1 does the same near 1200 iterations. We time. Due to the new updating scheme (sample to sample) a new
used a fixed number of iterations per kind of algorithm in order evaluation function is proposed as
to compare the execution time. Also, it is important to point out
2 2
that after several experiments, neither increasing the number of 1
J(k) = [(T̂ i(k) − Ti(k) ) + (Ĥ i(k) − Hi(k) ) ]. (27)
particles/individuals nor augmenting the number of iterations per 2
algorithm achieve better results. Taking into account the above
The algorithms are implemented in a 72 h lapse, starting on April
results, the best representative algorithms, ICFPSO and DE current
7, 2014, at 12:45 p.m. In order to prove the adaptation capability of
to best/1, are reprogrammed in order to construct an online version
the optimized model, the experimental conditions were changed:
of each one. The next subsection presents the results for the online
the fans were turned on and the windows kept full open during all
optimization schedule.
the execution. Fig. 8 shows the climatic data measured during this
new experiment.
6.3. Online implementation As a final adaptation, calibration parameters from the respective
offline version are used, with the addition of an initial condition for
The previous subsections show the results of the selected PSO both algorithms, which generates the first set of particles in a neigh-
and DE algorithms, executed in an offline stage, where the evalu- borhood of ±10% from each term of the best solution found in the
ation process occurs in each particle in each iteration using all the offline stage. Figs. 9 and 10 show the results for online ICFPSO and
samples of the data base presented in Fig. 2. The obtained results online DE current to best/1, respectively, after the 72 h of execution.
place as the algorithms with the best performance, in first instance, Adding a code line with an accumulator of the approximation
the DE current to best/1, followed by the ICFPSO. In this section, error, experimentation delivers a total MSE of 5.64 units for ICFPSO,
those algorithms are re-evaluated in an online scheme through a a great performance improvement in comparison with the offline
real-time execution obtained by using the properties of LabViewTM version, and 9.4237 units for DE current to best/1, which represents
RT and the CompactRIO programming tools. The execution scheme a non-significant improvement of the solution. It is recalled that for
A. Pérez-González et al. / Applied Soft Computing 62 (2018) 86–100 97

Fig. 8. Climate behavior during the online ICFPSO and DE current to best/1 execution.

Fig. 9. Online ICFPSO graphic results.

the offline scenario, the DE algorithm shows a better performance out that parameters found by the online algorithms were omitted
in parameter identification, as well as in execution time, but in an from this paper, due to the variations presented in consequence of
online scenario the PSO shows a better adaptation capability, even the sample to sample update scheme.
with a drastically reduced population. Also, it is important to point
98 A. Pérez-González et al. / Applied Soft Computing 62 (2018) 86–100

Fig. 10. Online DE current to best/1 graphic results.

Experimental results allow to define completely the identifi- As future work, the proposed methodology will be proven again
cation scheme of better performance: first, a pre-identification in the final building stage of the greenhouse prototype and in pres-
stage using DE current-to-best/1 algorithm, which has an execu- ence of crops, in offline and online mode, while the optimized
tion time of 18 h, 6 min and 19.8 s for a database of 24 h; then, the model will be used to design control actions to manipulate the
obtained results are taken as initial conditions of an online refine- greenhouse internal temperature and relative humidity. Also, it is
ment scheme based on ICFPSO, with a time consumption of 52.4% intended to implement the methodology in other greenhouses, in
of the time between samples, using a sampling time of 1 s. order to prove its adaptation capability.
Finally, Table 8 presents a summary of several approaches avail-
able in the current literature related to greenhouse modeling and Acknowledgments
its parameter identification, including comments for each one, that
allow to establish comparison points between them and the pro- We sincerely thank to CONACyT for the economic incentives
posed methodology, including of course the results obtained in this provided during this investigation, to CINVESTAV for the provided
paper. resources for the greenhouse prototype construction and instru-
mentation, and to the researchers involved in the development of
this project.
7. Conclusions

In this paper, a generic mathematical model for a greenhouse References


has been taken from the literature, adapting it to the physical
[1] R. Castañeda-Miranda, E. Ventura-Ramos Jr., R. Peniche-Vera, G. Herrera-Ruiz,
conditions of a real prototype, and then its parameters were iden- Analysis and simulation of a greenhouse physical model under weather
tified in order to improve the correspondence of that model to the conditions of the central region of México, Agrociencia 41 (3) (2007) 317–335.
dynamic behavior of the prototype. The identification process has [3] X. Blasco, M. Martínez, J. Herrero, C. Ramos, J. Sanchis, Model-based predictive
control of greenhouse climate for reducing energy and water consumption,
been divided in two stages, first, testing a set of PSO and DE algo- Comput. Electron. Agric. 55 (1) (2007) 49–70.
rithms whose properties are given by its calibration parameters, [4] R. Salazar, I. Lopez, A. Rojano, A neural network model to control greenhouse
and second, introducing the optimization scheme in a real-time environment, in: Sixth Mexican International Conference on Artificial
Intelligence-Special Session, MICAI 2007, IEEE, 2007, pp. 311–318.
execution using LabViewTM . The related code was validated in the [5] M. Nachidi, A. Benzaouia, F. Tadeo, Temperature and humidity control in
real prototype, using also real data records as online measurements. greenhouses using the Takagi–Sugeno fuzzy model, in: 2006 IEEE
The comparison of the behavior of the optimized model and the International Conference on Control Applications, 2006 IEEE International
Symposium on Intelligent Control, 2006 IEEE Computer Aided Control System
real system through the MSE calculus provides a practical criterion Design, IEEE, 2006, pp. 2150–2154.
to select the best performance algorithm to optimize a problem [6] C. Von Zabeltitz, Greenhouse structures, Ecosystems of the world (1999)
of this complexity, wherever appropriate calibration parameters 17–70.
[7] R.J. Van Ooteghem, Optimal Control Design for a Solar Greenhouse, vol. 3,
are selected. Also, the sample to sample adaptation of the selected
2010.
algorithms increases the already good behavior of the computa- [8] J.L. Iga, Modeling of the climate for a greenhouse in the north-east of México,
tional methods, allowing the optimization to be faster and better The International Federation of Automatic Control Proceedings of the 17th
World Congress (2008) 9558–9563.
in the online form than the offline one. It is important to emphasize
[9] K.R. Manohar, C. Igathinathane, Greenhouse Technology and Management, BS
that, based on the obtained results, we propose a methodology to Publications, 2007.
deal with a greenhouse identification problem based on the selec- [10] F.O. Ruiz Palacios, C.E. Cotrino Badillo, Identifying a greenhouse climate model
tion of a mathematical representation, and afterwards performing by using subspace methods, Ingeniería e Investigación 30 (2) (2010) 157–167.
[11] H.U. Frausto, J. Pieters, J. Deltour, Modelling greenhouse temperature by
an offline pre-identification using the DE current to best/1 algo- means of auto regressive models, Biosyst. Eng. 84 (2) (2003) 147–157.
rithm due to its low time consumption, and finally, performing an [12] A. Hasni, R. Taibi, B. Draoui, T. Boulard, Optimization of greenhouse climate
online identification stage using the ICFPSO, owing to its adaptation model parameters using particle swarm optimization and genetic algorithms,
Energy Proc. 6 (2011) 371–380.
capability in a sample to sample execution. Estimation results are [13] J. Coelho, P. de Moura Oliveira, J.B. Cunha, Greenhouse air temperature
superior in this work than in other papers reported in the literature, predictive control using the particle swarm optimisation algorithm, Comput.
where the online stage additionally grants the model a time variant Electron. Agric. 49 (3) (2005) 330–344.
[14] E. Cruz-Valeriano, O. Begovich, J. Ruiz-León, Modeling of a greenhouse using
parameter identification property. Furthermore, a control strategy particle swarm optimization, in: 2013 10th International Conference on
based on the availability of those parameters can be implemented Electrical Engineering, Computing Science and Automatic Control (CCE), IEEE,
as a sample to sample adaptive algorithm. 2013, pp. 268–273.
A. Pérez-González et al. / Applied Soft Computing 62 (2018) 86–100 99

Table 8
Comparison of the offline-online heuristic scheme against different methods.

Approach Comments

Acknowledgement model with physical adaptation [8] This approach only considers the temperature inside a greenhouse prototype. It
reports a maximum deviation on the model adaptation of 7.09 ◦ C and an average
deviation of 1.21 ◦ C. Our approach reaches a maximum deviation for the temperature
of 2.47 ◦ C and an average deviation of 1.87 × 10−3 ◦ C.
Subspace methods [10] It includes the estimation of foliage temperature, omitted in this work due to the
building stage of the prototype. This reference reports a combined MSE for
temperature and relative humidity of 6.49 units, 0.85 units above the result of our
paper.
Means of auto-regressive models [11] It presents only the estimation of temperature for a greenhouse, but includes a
complete seasonal analysis of the adaptation of the proposed model. The parameter
identification only has regressive value without a physical interpretation. The model
presents falls on the goodness of fit below 70% due to the season change. This issue can
be corrected with the online sample-to-sample schedule proposed in our work.
PSO and GAs approach [12] It only reports the temperature estimation, with a MSE for PSO of 0.1280 units, and for
GA of 0.1341 units, but considering only 161 samples. In the experiments of our work,
the MSE seems to be bigger, but it is important to highlight that the total number of
samples is 518400. The graphical comparison of the behavior of the model estimations
shows a better fit in our research.
Classical PSO [14] It presents the estimations for temperature and relative humidity, but reports high
displacements between the estimations and measurements when the real prototype
begins to lose thermal energy (at sunset). These displacements are impossible to be
compensated by the selected PSO in [14].
PSO and GAs approach [15] This work is the predecessor of [12]; it only presents the temperature estimations, but
includes a PSO form with variations in the inertia value. The number of estimated
parameters is only 4, against 17 that are estimated in our work. It does not present an
online scheme. The best results present a maximum displacement of temperature
estimation higher than 5 ◦ C, against 2.47 ◦ C as the maximum deviation presented in
our work.
Online PSO approach [17] This work introduces the use of the online heuristic algorithm. It omits the use of
another kind of heuristics, and because of that, the computational time for the
pre-identification stage is considerable bigger than the best performance reached in
our work.
Multi-objective evolutionary algorithms [30] This research introduces the Pareto analysis in the objective function, and implements
an identification of 15 different physical parameters, reporting an excellent result with
0.845 ◦ C and 2.564% of displacement between the measurements and the estimations.
Nevertheless, the model identification shows an important displacement in the
relative humidity estimations at the moment of saturation, with differences of almost
10%. As in [11], this issue can be corrected by the use of an online scheme.
Robust identification using evolutionary algorithms [31] The ε − GA algorithm, which characterizes the Feasible Parameter Set by means of a
discrete set of models that are well distributed, is presented as the methodology of
solution. Multi-model optimization confers to this research a high level of robustness,
but, in the same way as its predecessor [30], the approach presents high
displacements in the humidity saturation area.
Extreme learning approach [32] It presents a set of algorithms used as the training group for neural networks, and
introduces the wireless instrumentation. The model implementation is based on
Black-box approach, and presents an average fit on the prediction between 97% and
82% against 93% of fit as the worst value reached in our work. A value of 3.1305 units
of combined MSE are presented as the best result reached in [32], which seems to be a
better result. Calculation of MSE is lower due to the low number of samples
implemented.
Least squares estimation [33] This paper presents both, identification and control for a greenhouse prototype. The
combined MSE obtained using the LSE is 0.09201 units for 14400 samples, also, the
method omits the physical meaning of parameters. Graphical results are more
representative and allow to note high displacements between estimations and
measures in the extreme values of temperature and relative humidity.
Means of neural networks [34] A NNARX system is proposed as model identifier in this paper, and it only considers the
estimations of temperature. Graphical results show a good match between estimations
and measurements, and the goodness of fit is reported with values above 75%.
Nevertheless, the method omits a physical interpretation for the identification results.
Discrete events characterization [35] This research introduces a simplification for the acknowledgement model, presenting
it as a decomposition based on a set of simpler submodels. In graphical results, it is
possible to observe that submodel transitions produce a high-noise estimation
behavior.
Fuzzy clustering approach [36] This work presents a decomposition on linear models for a greenhouse representation.
The results reach a high performance estimating the temperature and relative
humidity, but physical meaning of the estimated parameters is omitted. The graphical
results show presence of disturbances in few points of the evolution of the
experiment, which presents displacements above 8 ◦ C and 15% for temperature and
relative humidity, respectively.
ART2 classifier [37] It presents a decomposition using local models obtained with an Elman Neural
Network technique from a greenhouse data base, and also introduces a supervisory
controller based on neural approach. The work is centered in the controller results,
presenting few statements about the model identification. Nevertheless, graphical
results allow to observe big displacements in temperature and relative humidity
estimations.
100 A. Pérez-González et al. / Applied Soft Computing 62 (2018) 86–100

[15] A. Hasni, B. Draoui, M. Latfaoui, T. Boulard, Identification of natural ventilation r: mix ratio
parameters in a greenhouse with continuous roof vents, using a PSO and GAs, S: exchange surface between components (m2 )
Sens. Transducers 119 (8) (2010) 182. s: windows opening surface (m2 )
[16] R. Avila-Miranda, O. Begovich, J. Ruiz-León, An optimal and intelligent control s0 : leakage surface (m2 )
strategy to ventilate a greenhouse, in: 2013 IEEE Congress on Evolutionary T* : air temperature (◦ C)
Computation (CEC), IEEE, 2013, pp. 779–782. Tm : thermic mass temperature (◦ C)
[17] A. Pérez-González, O. Begovich, J. Ruiz-León, Modeling of a greenhouse
V: wind speed (m/s)
prototype using PSO algorithm based on a LabView application, in: 2014 11th
v: greenhouse volume (m3 )
International Conference on Electrical Engineering, Computing Science and
˛: coefficient of absorption of solar radiation due to air
Automatic Control (CCE), IEEE, 2014, pp. 1–6.
[18] C. Arredondo, J. Hernández, G. Gordillo, et al., Development of equipment for ˇ: coefficient of absorption of solar radiation due to mass
monitoring PV power plants, using virtual instrumentation, in: Conference : psychometric constant (hPa/K)
Record of the 2006 IEEE 4th World Conference on Photovoltaic Energy t: sampling time (s)
Conversion, Vol. 2, IEEE, 2006, pp. 2367–2370. , , : system mediation coefficients
[19] E. Koutroulis, K. Kalaitzakis, Development of an integrated data-acquisition : air density (kg/m3 )
system for renewable energy sources systems monitoring, Renew. Energy 28 : characteristic time (s)
(1) (2003) 139–152.   : coating transitivity coefficient
[20] J.V. Iribarne, W.L. Godson, Atmospheric Thermodynamics, vol. 6, Springer ϕi : humidification system power (W/m2 )
Science & Business Media, 2012. ϕl : cooling system power (W/m2 )
[21] R.C. Eberhart, J. Kennedy, A new optimizer using particle swarm theory, in: *i : internal dynamic
Proceedings of the Sixth International Symposium on Micro Machine and
*e : external dynamic
Human Science, Vol. 1, New York, NY, 1995, pp. 39–43.
[22] Q. Bai, Analysis of particle swarm optimization algorithm, Comput. Inf. Sci. 3
(1) (2010) 180.
[23] M.E.H. Pedersen, Tuning & Simplifying Heuristical Optimization (Ph.D. thesis), Alfonso Pérez González graduated from University of
University of Southampton, 2010. Guadalajara with the engineer in electronics and com-
[24] M.E.H. Pedersen, A.J. Chipperfield, Simplifying particle swarm optimization, munications degree in 2012. Academic Excellence Award
App. Soft Comput. 10 (2) (2010) 618–628. from the National Association of Faculties and Schools of
[25] M. Clerc, J. Kennedy, The particle swarm-explosion, stability, and convergence Engineering in 2012. M.Sc. degree obtained in the Center
in a multidimensional complex space, IEEE Trans. Evol. Comput. 6 (1) (2002) for Research and Advanced Studies (CINVESTAV) Campus
58–73. Guadalajara in the field of electric engineering in 2014.
[26] R. Storn, K. Price, Differential evolution – a simple and efficient heuristic for Academic studies alternated with industrial experience
global optimization over continuous spaces, J. of Glob. Optim. 11 (4) (1997) since 2010 in various companies in areas of electronic
341–359. diagnosis and repair, automation, programming, fault
[27] K.V. Price, An introduction to differential evolution, in: New Ideas in analysis, and electromechanical R + D. Currently conduct-
Optimization, McGraw-Hill Ltd, UK, 1999, pp. 79–108. ing doctoral studies at CINVESTAV Campus Guadalajara in
[28] K. Price, R.M. Storn, J.A. Lampinen, Differential Evolution: A Practical the field of metaheuristic optimization.
Approach to Global Optimization, Springer Science & Business Media, 2006.
[29] M.E.H. Pedersen, Good Parameters for Particle Swarm Optimization. Tech.
Ofelia Begovich Mendoza was born in México City, Méx-
Rep. HL1001, Hvass Lab., Copenhagen, Denmark, 2010.
ico. She obtained the M.Sc. in electrical engineering from
[30] J. Herrero, X. Blasco, M. Martínez, C. Ramos, J. Sanchis, Non-linear robust
CINVESTAV, México City, in 1987, and the Ph.D. degree
identification of a greenhouse model using multi-objective evolutionary
in electrical engineering from the University of Rennes
algorithms, Biosyst. Eng. 98 (3) (2007) 335–346.
I, Rennes, France in 1992. Since 1992, she is a professor
[31] J. Herrero, X. Blasco, M. Martínez, C. Ramos, J. Sanchis, Robust identification of
of electrical engineering of graduate program in CINVES-
non-linear greenhouse model using evolutionary algorithms, Control Eng.
TAV, Guadalajara, México. Her major research field is the
Pract. 16 (5) (2008) 515–530.
control of processes, leak diagnosis, control and monitor-
[32] Q. Liu, D. Jin, J. Shen, Z. Fu, N. Linge, A WSN-based prediction model of
ing of greenhouses and heuristic optimization applied to
microclimate in a greenhouse using extreme learning approaches, in: 2016
parameter estimation.
18th International Conference on Advanced Communication Technology
(ICACT), IEEE, 2016, pp. 730–735.
[33] N. Bennis, J. Duplaix, G. Enéa, M. Haloua, H. Youlal, Greenhouse climate
modelling and robust control, Comput. Electron. Agric. 61 (2) (2008) 96–107.
[34] H.U. Frausto, J.G. Pieters, Modelling greenhouse temperature using system
identification by means of neural networks, Neurocomputing 56 (2004)
423–428.
[35] L.H. Rajaoarisoa, N.K. M’Sirdi, E. Boukas, J.F. Balmat, J. Duplaix, Greenhouse Javier Ruiz León received the M.Sc. degree from CINVES-
climate modeling: observability and identification for supervision, in: Proc. of TAV, México, in 1992, and the Ph.D. degree from the Czech
the 5th International Conference on Informatics in Control, Automation and Technical University in Prague, Czech Republic, in 1996.
Robotics, ICINCO, 2008, Vol. 8. He is currently working at CINVESTAV Campus Guadala-
[36] A. Errahmani, M. Benyakhlef, I. Boumhidi, Greenhouse model identification jara, México. His research interests include linear system
based on fuzzy clustering approach, ICGST-ACSE J. 9 (2009) 23–27. theory and hybrid systems.
[37] F. Fourati, Multiple neural control of a greenhouse, Neurocomputing 139
(2014) 138–144.

Glossary

A: saturation gain due to transpiration (g/m3 )


Al : windows natural ventilation factor
B: transpiration coefficient (W/m2 hPa)
C: natural ventilation factor due to air convection
Cl : water steam thermic capacity (1/kg K)
Cm : heavy components thermic capacity (1/kg K)
Cp : air thermic capacity (1/kg K)
d0 : leakage factor (m3 /s)
H* : relative humidity (%)
h: thermic exchange convection coefficient (W/m2 K)
K: thermal loss coefficient (W/m2 K)
Kl : latent heat transfer coefficient (W/m2 hPa)
Ks : sensible heat transfer coefficient (W/m2 hPa)
P* : pressure (hPa)
P(T* ): saturation pressure (hPa)
Qair : air thermic charge (W/m2 )
Qsol : soil thermic charge (W/m2 )
Rg : solar radiation (W/m2 )

Anda mungkin juga menyukai