Professional Documents
Culture Documents
1 s2.0 S0308521X22000245 Main
1 s2.0 S0308521X22000245 Main
Agricultural Systems
journal homepage: www.elsevier.com/locate/agsy
Review
H I G H L I G H T S G R A P H I C A L A B S T R A C T
A R T I C L E I N F O A B S T R A C T
Editor: Mark van Wijk CONTEXT: Process-based greenhouse climate models are valuable tools for the analysis and design of greenhouse
systems. A growing number of greenhouse models are published in recent years, making it difficult to identify
Keywords: which components are shared across models, which are new developments, and what are the objectives,
Review strengths and weaknesses of each model.
Greenhouse model
OBJECTIVE: We present an overview of the current state of greenhouse modelling by analyzing studies published
Climate model
between 2018 and 2020. This analysis helps identify the key processes considered in process-based greenhouse
Crop model
Model evaluation models, and the common approaches used to model them. Moreover, we outline how greenhouse models differ in
Model validation terms of their objectives, complexity, accuracy, and transparency.
METHODS: We describe a general structure of process-based greenhouse climate models, including a range of
common approaches for describing the various model components. We analyze recently published models with
respect to this structure, as well as their intended purposes, greenhouse systems they represent, equipment
included, and crops considered. We present a model inheritance chart, outlining the origins of contemporary
models, and showing which were built on previous works. We compare model validation studies and show the
various types of datasets and metrics used for validation.
* Corresponding author at: Wageningen UR Greenhouse Horticulture, PO Box 644, 6700 AP Wageningen, the Netherlands.
E-mail address: [email protected] (D. Katzin).
https://doi.org/10.1016/j.agsy.2022.103388
Received 17 September 2021; Received in revised form 21 January 2022; Accepted 4 February 2022
Available online 24 February 2022
0308-521X/© 2022 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
D. Katzin et al. Agricultural Systems 198 (2022) 103388
RESULTS AND CONCLUSIONS: The analysis highlights the range of objectives and approaches prevalent in
greenhouse modelling, and shows that despite the large variation in model design and complexity, considerable
overlap exists. Some possible reasons for the abundance of models include a lack of transparency and code
availability; a belief that model development is in itself a valuable research goal; a preference for simple models
in control-oriented studies; and a difference in the time scales considered. Approaches to model validation vary
considerably, making it difficult to compare models or assess if they serve their intended purposes. We suggest
that increased transparency and availability of source code will promote model reuse and extension, and that
shared datasets and evaluation benchmarks will facilitate model evaluation and comparison.
SIGNIFICANCE: This study highlights several issues that should be considered in greenhouse model selection and
development. Developers of new models can use the decomposition provided in order to present their models and
facilitate extension and reuse. Developers are encouraged to reflect on and explicitly state their model's range of
suitability, complexity, validity, and transparency. Lastly, we highlight several steps that could be taken by the
greenhouse modelling community in order to advance the field as a whole.
2
D. Katzin et al. Agricultural Systems 198 (2022) 103388
Fig. 1. Scheme of the greenhouse system. Control decisions are based on the outdoor weather, the indoor climate, and the crop status. The indoor climate is
influenced by the outdoor weather, the controls, and the crop. The crop is influenced by the indoor climate. Some outputs are yield (which depends on the crop), costs
and energy use (which depend on the controls). In this study we focus on the indoor climate and the processes influencing it.
3
D. Katzin et al. Agricultural Systems 198 (2022) 103388
grey-box) (Thornley and France, 2007). The distinction between the two existence, we also survey how this component was considered whenever
categories, however, is not always clear, and they are better viewed as it was included.
two edges on a spectrum (Keating and Thorburn, 2018). Descriptive
models describe systems using equations based on mathematical or 2.2. Objectives of process-based greenhouse climate modelling
statistical grounds, regardless of underlying principles. In contrast,
process-based models aim to provide an understanding or explanation of As mentioned earlier, process-based models are designed with the
the system being investigated, typically by combining two levels of intention that they provide insights that lie outside the domain of
description, with a lower level describing observed scientific phenom knowledge and data that was used in their development (Duncan, 1975;
ena, and a higher level describing emergent properties based on these Keating and Thorburn, 2018). The objectives and purposes of process-
phenomena (Thornley and France, 2007). The expectation is that based greenhouse climate models can be classified into four cate
process-based models generalize better than descriptive models to
gories: systems analysis, exploratory modelling, model-based control,
conditions outside the data range on which they have been developed and model-assisted design.
and validated. Thus, they can potentially provide insights that apply
In systems analysis, a model is used to better understand or describe
outside the limits of the system on which they were designed, predicting a particular greenhouse system. This approach often stems from a sci
the results of a range of “what-if” scenarios (Duncan, 1975; Keating and
entific, curiosity-driven, exploratory approach. Once the model is
Thorburn, 2018). adequately described and understood, it can be used for other purposes
One class of process-based greenhouse models treats the greenhouse
such as model-based design and control. In systems analysis, methods
air as a “perfectly stirred tank” (Roy et al., 2002). In this approach the such as sensitivity analysis can be used to reveal which components have
greenhouse air is treated as a uniform entity, where spatial variability is
a strong influence on the system. Here, care should be taken to be aware
ignored and representative values of, e.g., air temperature are used. In on what is actually being analyzed: the real-world system represented by
some cases, the air is divided into compartments such as the air above
the model, or the model itself, as a sensitivity analysis can uncover in
and below a thermal screen (see example below), but still each sights regarding both (e.g., Van Henten, 2003).
compartment is assumed to be perfectly stirred. Furthermore, under this
In exploratory modelling (also termed scenario analysis), the model
approach the greenhouse is often assumed (sometimes implicitly) to be is used to predict the results of untested scenarios. This analysis can
infinitely large (e.g., De Zwart, 1993). One consequence of this approach
point out directions for solving a problem or be used to narrow down a
is that the air is assumed not to be influenced by the side walls of the list of possible strategies or solutions which can then be tested in prac
greenhouse. tice. For example, De Zwart (1996) used a model to find the most
Another class of models uses computational fluid dynamics (CFD) to promising energy saving methods out of a predefined list.
describe the movement of air within the space of the greenhouse Model-based control uses models to apply methods such as model
(Boulard et al., 2002). This method, which is considerably more complex predictive control or optimal control on the greenhouse climate (e.g.,
and computationally intensive than the perfectly stirred tank approach, Katzin et al., 2020a; Kuijpers et al., 2021; Tap, 2000; Van Henten, 2003;
allows to describe heterogenous attributes of the greenhouse air and Van Henten et al., 1997; Van Ooteghem, 2007; Van Straten et al., 2010).
their change through space and time. A review on the possibilities and While this line of research generates meaningful insights regarding
challenges of CFD in greenhouse modelling was given by Norton et al. greenhouse climate control, it is rarely realized in commercial green
(2007), and more recent advances were listed by Choab et al. (2019). house practice (Van Beveren et al., 2015b).
Nevertheless, the heavy computational requirements of CFD models still Lastly, model-assisted design is a form of exploratory modelling used
limit their applicability, and a middle ground may be found by for the design of greenhouse systems. It may include scenario analysis or
combining them with perfectly stirred tank approaches (Piscia et al., more sophisticated methods. For example, Vanthoor (2011) presented a
2015). model and an optimization method which was used to find an optimal
Some greenhouse models are developed using building energy design (based on model predictions) for a given location and situation.
simulation programs such as EnergyPlus or TRNSYS (Choab et al.,
2019). These programs were designed to simulate the energy demand of
buildings, and considerable modifications are needed to correctly apply 2.3. General structure of process-based greenhouse climate models
them for greenhouses (Ahamed et al., 2020). In these platforms, models
are constructed using pre-existing components available within the In this section, we describe a general structure that is common in all
simulation program. While this could facilitate model development, it process-based greenhouse climate models. At the same time, we outline
reduces model transparency, since understanding the inner workings of the range of different approaches that are found between models. This
the model requires considerable knowledge of the simulation program section summarizes observations from several sources describing
that was used for its development. greenhouse models, with a wide range of complexity (De Zwart, 1996;
In this study, we focus on process-based, perfectly stirred tank Stanghellini et al., 2019; Van Henten, 1994; Van Straten et al., 2010).
models of the greenhouse climate. This means that we focus on the in The indoor climate may be described by one or more of the following
door climate and the processes that influence it (Fig. 1). An essential attributes: temperature, humidity, CO2 concentration, and light. A
component of the indoor climate is the air, but other components (e.g., general way to model the temperature, humidity, and CO2 concentration
crop temperature, floor temperature) may also be included. Models that is by considering balances: an energy balance, a water vapor balance,
describe exclusively the control system (e.g., the boiler, cogenerator, and a CO2 balance, but not all greenhouse climate models describe all
heat storage) or exclusively the crop (yield models), are outside the three balances: some focus only on energy, or only on energy and water.
scope of this study. Nevertheless, processes that influence the indoor For each of these balances, incoming and outgoing flows are identified,
climate (including crop processes such as photosynthesis and transpi and the difference between the incoming and outgoing flows is the net
ration) are reviewed here. Since crop yield is the most important change in each attribute. A set of equations that describes these net
component of the greenhouse system, and in fact, the reason for its changes is:
4
D. Katzin et al. Agricultural Systems 198 (2022) 103388
Fig. 2. Scheme of the Van Henten model (Eq. 1, Eq. 2) (Van Henten, 1994, 2003), based on the Forrester diagram conventions (Forrester, 1961; see also Haefner,
2005). Mass and energy flows are indicated by solid lines, information flows are indicated by dashed lines (see legend). Valves placed over solid lines indicate rate
equations that influence the rate of the flow passing through the valve. Information flows indicate the influence of an input or a state on a rate of flow. For example,
photosynthesis is influenced by solar radiation, crop dry weight, indoor air temperature and indoor CO2 concentration. Photosynthesis in turn influences the rate of
CO2 flow from the indoor air to the crop. Arrows indicate which objects act as sources, sinks, or both. For example, the outdoor air may be a source or a sink of water
vapor, but irrigation is only a source.
5
D. Katzin et al. Agricultural Systems 198 (2022) 103388
Fig. 3. Scheme of the De Zwart model (De Zwart, 1996) with its mass and energy flows (Eq. 1, Eq. 3). Influences on ventilation rate and air flow through screen are
depicted below; their influence on other flows are indicated by valves labelled by V and S, respectively. See Fig. 2 for legend and further explanation of the diagram
conventions. The photosynthesis rate influences the flow of CO2 from the indoor air to the crop. CO2 absorbed by the crop is an output of the greenhouse model which
is in turn used as an input for a crop model. To simplify the figure, exchanges of latent heat QLatent are shown together with the accompanying change of phase of
water WTrans, WCond, or WEvap. These processes simultaneously influence both the energy balance (a temperature of an object) as well as the water vapor balance (a
vapor concentration of an object). For example, condensation of vapor from the indoor air onto the screen influences both the indoor vapor concentration (WCond) as
well as the screen temperature (QLatent).
6
D. Katzin et al. Agricultural Systems 198 (2022) 103388
dE )
= QSun + QHeat + QLamp − QVent − QLatent − QCon − QFIR − QCool (J m− 2 s− 1
= W m− 2
dt
dMW )
= WTrans + WEvap + WHum − WCond − WVent − WDehum (kg {water vapor} m− 2 s− 1
(1)
dt
dMC )
= CInj − CPhot − CVent (kg {CO2 } m− 2 s− 1
dt
short wave and/or long wave radiation may be used as an input for a
Here, each line represents a balance. The first line represents the transpiration model (Katsoulas and Stanghellini, 2019). The amount of
energy balance, where dE dt is the net change of energy in the greenhouse,
available light in the greenhouse is a sum of the light originating from
with t representing a time unit (in this case seconds), and E expressed in the sun and from lamps. Light from the sun inside the greenhouse is
J m− 2. Expressions on the right-hand side of the equation are net flows typically a product of the solar radiation outside the greenhouse and the
which may be positive (adding energy, i.e., heating the greenhouse), greenhouse transmissivity, a unitless factor which describes what frac
negative (cooling the greenhouse), or zero. The sign in front of each tion of the outdoor radiation penetrates into the greenhouse. Light from
expression in Eq. 1 indicates what is the typical direction of each flow: the lamps is typically a factor of the energy provided to the lamps and
the typical incoming energy flows are QSun, heating from the sun; QHeat, the lamp's photosynthetic photon efficacy (PPE), which is expressed in
heat from the heating system; and QLamps, heat emitted by lamps. The μmol of photons of PAR light per J of energy input (Nelson and Bugbee,
typical outgoing energy flows are QVent, exchange of air through venti 2014).
lation; QLatent, conversion from sensible to latent heat; QCon, convective
and conductive exchanges with the outside; QFIR, thermal (far infrared) 2.3.1. Example1: The Van Henten model
radiation; and QCool, heat extracted by cooling mechanisms. A concrete example is the differential equations based model of Van
The second line represents the water vapor balance, with dM Henten (1994, 2003), summarized here and in Fig. 2 using notation from
dt the net
W
Eq. 1:
change of water vapor mass in the greenhouse. The typical incoming
flows are WTrans, crop transpiration; WEvap, water evaporation from the dXT 1 (◦ C s-1)
= (QSun + QHeat − QVent − QCon )
soil or other surfaces; and WHum, which includes humidity added by dt ccap,q
dXh 1 (kg {water vapor} m-2 s-1)
control mechanisms such as fogging or pad and fan cooling. The typical = (WTrans − WVent )
dt ccap,h
outgoing flows are WCond, condensation of vapor on cold surfaces; WVent, dXc 1 ( ) (kg {CO2} m-2 s-1) (2)
vapor exchange through ventilation; and WDehum, vapor extracted by = CInj − CPhot − CVent
dt ccap,c
dehumidification mechanisms. Water vapor released by crop transpi dXd
= f(XT , XC , Xd , Vrad ) (kg {dry weight} m-2 s-1)
ration originates from irrigation, which is typically assumed to be sup dt
CPhot = g(XT, Xc, Xd, Vrad) (kg {CO2} m-2 s-1)
plied in a sufficient rate such that the water availability does not reduce WTrans = h(XT, Xh, Xd) (kg {water vapor} m-2 s-1)
transpiration.
The third line represents the CO2 balance, where dMdt is the net change
C
of CO2 mass in the air. The typical incoming flow is CInj, enrichment of This model is composed of 4 states described by 4 differential
the air by CO2 injection; the typical outgoing flows are CPhot, crop net equations. Three of these states correspond to the balances of Eq. 1: the
photosynthesis; and CVent, CO2 exchange through ventilation. indoor temperature XT (◦ C), the indoor vapor concentration Xh (kg
Some of the flows above may act both as incoming and outgoing {water vapor} m− 3), and the indoor CO2 concentration Xc (kg {CO2}
flows. For example, when CVent is positive it represents an outgoing flow, m− 3), defined by equations corresponding to the balances E, MW, and MC
losses of CO2 from the system through ventilation. This is the common of Eq. 1. The parameters ccap,q, ccap,h, ccap,c and the functions f(XT, XC, Xd,
case when the indoor CO2 concentration is higher than the outdoor. Vrad), g(XT, XC, Xd, Vrad), and h(XT, Xh, Xd) are described in Van Henten
However, CVent may also be negative, for instance, if the indoor CO2 (1994, 2003). Several components mentioned in Eq. 1 are neglected in
concentration is lower than the outdoor. In this case the expression this model. At the same time, the model describes a state Xd representing
− CVent in Eq. 1 will be positive and represent an incoming flow of CO2 to the dry weight of the crop in the greenhouse (kg {dry weight} m− 2). This
the system. state, governed by the equation f whose definition is excluded here,
While Eq. 1 describes the greenhouse balances as differential equa provides additional information about the greenhouse system but it is
tions, not all process-based greenhouse climate models describe or not part of the climate balances. Nevertheless, the crop dry weight state
simulate these balances in this way. For example, the use of discrete- Xd influences the photosynthesis and transpiration flows CPhot and WTrans
time difference equations is common (Lopez-Cruz et al., 2018). (given here as functions g and h), which are part of the greenhouse
Furthermore, as can be seen in Fig. 1, many of the greenhouse compo climate system.
nents are interdependent: for example, crop transpiration and photo
synthesis are influenced by the indoor climate, which is in turn 2.3.2. Example 2: The De Zwart model (KASPRO)
influenced by crop transpiration and photosynthesis. Therefore, the In order to illustrate the range of models that are represented by Eq.
model may need to be solved iteratively. Another approach is to assume 1, another example (De Zwart, 1996) is given in Fig. 3. The De Zwart
dMW dMC model (also known as KASPRO), developed around the same time and
that the entire system is in steady state, i.e., that dE
dt = 0, dt = 0, dt = place as the Van Henten model, is remarkably more elaborate. It in
0. Using this approach allows to model all but one of the components in
cludes nearly all components listed in Eq. 1, with several of them further
each line of Eq. 1, and calculate the last component based on the steady
decomposed to smaller subcomponents. For example, QSun is divided
state assumption.
into diffuse and direct radiation from the sun and is composed of solar
Light is an important attribute of the indoor climate of the green
radiation heating the greenhouse floor, air, crop, and cover. Using the
house. In particular, the amount of photosynthetically active radiation
notation of the De Zwart model, the various components of the energy
(PAR) in the greenhouse may be used as an input to a crop model in
order to estimate crop growth and photosynthesis, and the amount of
7
D. Katzin et al. Agricultural Systems 198 (2022) 103388
balance are defined by: approaches is possible for modelling the greenhouse climate. These
QSun = PSunCov + PSunAir + PVISCan + PNIRCan + PVISFlr + PNIRFlr (W m-2) approaches can be classified on a spectrum between “simple” and
QHeat = HBoilUpp + HBoilLow (W m-2) “complex”. Complex models include a larger number of processes and
QLamp = PAluAir (W m-2) objects, and use mechanistic descriptions of processes that involve
QVent = HTopOut + HAirOut (W m-2) (3)
multiple influencing variables or inputs. Simpler models neglect some
QLatent = LCanAir + LScrTop − LAirScr − LTopCov − LAirCov (W m-2)
QCon = HCovOut (W m-2) processes, use fewer objects, and summarize phenomena with descrip
QFIR = RCovSky (W m-2) tive functions, while maintaining an overall process-based model
structure. Objects in this context are entities that are described by var
iables: objects of the energy balance are described by their temperature,
Here, P is net shortwave radiation, H is net convection or conduction, objects of the water vapor balance are described by their vapor con
L is net latent energy, and R is net thermal radiation. The subscripts centration, and objects of the CO2 balance are described by their CO2
indicate the origin and target of the net energy flow: for example, HCovOut concentration. For example, the Van Henten energy balance includes
represents net energy exchange by convection from the greenhouse only two objects: indoor and outdoor air, and only 4 processes: solar
cover to the outside air. When HCovOut is positive, the net energy transfer radiation, heating pipes, convection and ventilation (Fig. 2, Eq. 2). The
is from the cover to the outside air, cooling the greenhouse; when HCovOut De Zwart energy balance has 17 objects, including soil layers and the sky
is negative, the net energy transfer is from the outside air to the cover, temperature, and processes not considered by Van Henten such as
heating the greenhouse.. Note that PVISCan and PNIRCan are denoted by a thermal radiation (FIR) and conversion to latent heat (Fig. 3, Eq. 3, Eq.
single line in Fig. 3, and similarly for PVISFlr and PNIRFlr. Similar equations 4).
as in Eq. 3 can be constructed for the water and CO2 balances. Besides the objects and processes included, models also vary in how
The De Zwart model is made up of 14 states that are defined using each process is described. Table 1 lists simple and complex approaches
ordinary differential equations. These states are: the vapor pressure of that are used to describe some of the processes in Eq. 1. The following
the air below the screen, the CO2 concentration of the air below the subsections provide further detail.
screen, and the temperatures of the cover, air below the screen, crop,
upper heating net, lower heating net, floor, and 6 soil layers. Four 2.4.1. Solar radiation and energy from lamps
additional variables are calculated using algebraic equations. These The heating input from the sun can be described as QSun = aSunISun
variables are: the vapor pressure of the air above the screen, the CO2 (W m− 2) with ISun (W m− 2) representing solar radiation from the sun and
concentration of the air above the screen, the temperature of air above aSun (− ) the fraction of global radiation that contributes to heating the
the screen, and the temperature of the screen (see Eq. 4, below). As in greenhouse. Solar radiation ISun is typically given as an input to the
the case of the Van Henten model, the states serve two purposes: first, model. This input may be a single value representing global radiation, or
they add detail to the simulation, providing descriptions for the tem two values differentiating between direct and diffuse radiation. Solar
peratures of the cover, air, canopy, etc. Second, these details are used to radiation can further be decomposed into photosynthetically active ra
calculate the inflows and outflows of the general balance equations (Eq. diation (PAR), which is used by the crop model to calculate photosyn
3). For example, the net flow of thermal radiation QFIR is defined as the thesis, and other wavebands, such as near infrared radiation (NIR) or
thermal radiation from the cover to the sky RCovSky. This value depends ultraviolet (UV) radiation which contribute to heating but not to
on the cover temperature TCov, as will be illustrated in the next section. photosynthesis. The coefficient aSun may be assumed constant, or
To give more concrete detail, some of the equations concerning the depend on the location of the sun in the sky and the amount of diffuse
energy balance in the De Zwart model are: and direct radiation. Note that the value of aSun can be different from
dTcov 1 ( ) (◦ C s-1)
= P + HTopCov + HAirCov + RFlrCov + RScrCov + RUppCov + RLowCov + RCanCov + LTopCov + LAirCov − HCovOut − RCovSky
dt ρcov cp,cov Vcov SunCov
0 = HScrTop + HAirTop − HTopCov − HTopOut (◦ C s-1)
0 = RFlrScr + LAirScr + RUppScr + RLowScr + HAirScr + RCanScr − HScrTop − LScrTop − RScrCov (◦ C s-1)
dTair 1 ( ) (◦ C s-1)
= P + PSunAir + HUppAir + HLowAir + HCanAir − HAirFlr − HAirTop − HAirScr − HAirOut − HAirCov
dt ρair cp,air Vair AluAir
dTcan 1 ( ) (◦ C s-1)
= RUppCan + RLowCan + PVISCan + PNIRCan − HCanAir − RCanCov − RCanScr − RCanFlr − LCanAir
dt capleaf LAI
(4)
dTflr 1 ( ) (◦ C s-1)
= R + RUppFlr + HAirFlr + RCanFlr + PVISFlr + PNIRFlr − HFlrSo1 − RFlrScr − RFlrCov
dt ρflr cp,flr Vflr LowFlr
dTupp 1 ( ) (◦ C s-1)
= H − RUppScr − RUppCov − HUppAir − RUppCan − RUppFlr
dt ρupp cp,upp Vupp BoilUpp
dTlow 1 (◦ C s-1)
= (H − RLowScr + RLowCov + HLowAir + RLowCan + RLowFlr )
dt ρlow cp,low Vlow BoilLow
dTso(i) 1 ( ) (◦ C s-1)
= HSo(i− 1)So(i) − HSo(i)So(i+1) i = 1, …6
dt thso(i) ρcp,soil
8
D. Katzin et al. Agricultural Systems 198 (2022) 103388
9
D. Katzin et al. Agricultural Systems 198 (2022) 103388
Table 2
Overview of studies related to process-based greenhouse climate modelling, 2018–2020. Category indicates the category of the study. Development is the development
status of the model used in the study. Greenhouse type: PE: polyethylene, CSG: Chinese solar greenhouse. Equipment: B: boiler, BS: blackout screen, C: cooling, CHP:
cogenerator, CO2: CO2 injection, DAH: direct air heating, FG: fogging, FN: fans, GP: grow-pipes, H: heating, HPS: high-pressure sodium lamps, HS: heat storage, HU:
humidification, L: lamps (no lamp type specified), LED: light-emitting diodes, MHL: metal halide lamps, PV: photovoltaic cells, P&F: pad and fan cooling, TS: thermal
screen. All models include ventilation and all CSG models include a thermal blanket and heat storage in a wall. Crop: None: no crop was present in the greenhouse. n/a:
information was not available. See Section 3.1 for more details.
Reference Category Purpose Greenhouse type Equipment Crop Development
Abbes et al. (2019) Control Develop a greenhouse model for North African context to PE tunnel – None Extension
help control microclimate
Ahamed et al. Exploratory Develop a model for designing and estimating heating CSG CO2, FN, H, L Tomato Extension
(2018a) demands of CSGs
Ahamed et al. Design Design an energy efficient greenhouse Double-layer PE CO2, FN, L, TS Tomato Reuse
(2018c)
Ahamed et al. Analysis Estimate heating demands of CSG CSG double layer CO2, FN, H, L Tomato Reuse
(2018d) PE
Ahamed et al. (2020) Analysis Develop a TRNSYS based model for CSGs CSG, glass cover CO2, FN, H, L n/a Translation
Alinejad et al. (2020) Design Asses an adjustable PV blind system for greenhouses Multi-span flat FN, H, P&F, PV, TS Rose New
arch PE
Baglivo et al. (2020) Analysis Develop a TRNSYS based model for greenhouses Venlo C, H, HPS Chrysanthemum New
Chen et al. (2019b) Design Design a CSG CSG – n/a New
Chen et al. (2019a) Design Evaluate placement of PV cells on a greenhouse roof Venlo PV Vriesea Extension
De Ridder et al. Calibration Propose a method for model calibration Venlo H, HPS, LED, TS Cucumber New
(2020)
Esmaeli and Design Optimize the design of a CSG CSG – Various New
Roshandel (2020)
Gharghory (2020) Calibration Use a deep network to predict greenhouse indoor climate n/a FG, H n/a Reuse
Golzar et al. (2018) Design Develop a greenhouse model that includes energy Venlo CO2, H, HU, MHL Tomato New
demand and yield in order to optimize greenhouse design
Golzar et al. (2019) Analysis Investigate the most important drivers for environmental Venlo CO2, H, HU, MHL Tomato Reuse
impacts of greenhouses
Hemming et al. Exploratory Compare greenhouse control strategies under different Venlo CO2, HPS, TS Cucumber Extension
(2019b) settings
Jomaa et al. (2019) Control Use fuzzy logic to control the greenhouse temperature n/a FG, H n/a Reuse
Katzin et al. (2020b) Exploratory Design a greenhouse model to predict the implications of Venlo BS, CO2, H, HPS, Tomato Extension
changing the greenhouse lighting system LED, TS
Lammari et al. Control Perform model calibration and setpoint tracking using Multi-span arch, FG, H Tomato Reuse
(2020) proportional integral sliding mode controllers PE
Ma et al. (2019) Control Predict the microclimate to achieve climate homogeneity Venlo HPS, P&F n/a New
with a conveyor belt system
Mohamed and Control Use an adaptive neuro fuzzy interface system to control a n/a FG, H None Reuse
Hameed (2018) greenhouse
Mohammadi et al. Analysis Develop a model for a semi-solar greenhouse CSG, glass cover – Cabbage New
(2020)
Pérez-González et al. Calibration Apply particle swarm optimization and differential Single span arch FG, H None Extension
(2018) evolution to parametrize a greenhouse model
Rasheed et al. (2019) Design Propose a reliable greenhouse model by using TRNSYS, Gambrel roof, H, TS None New
with a focus on thermal screens double PE
Righini et al. (2020) Exploratory Extend a greenhouse model by adding lamps, heat Venlo CO2, GP, H, HPS, Tomato Extension
harvesting and test how energy can be saved HS, LED, TS
Seginer et al. Control Evaluate the advantage of expanding the indoor climate Venlo B, CHP, HS Tomato Extension
(2020b) bounds
Sethi (2019) Design Develop a model for an asymmetric overlap roof shape PE FG, FN Tomato New
(AORS) greenhouse
Su et al. (2018) Control Apply optimal control based on adaptive dynamic Venlo DAH, F Tomato Extension
programming to save energy
Subin et al. (2020) Control Implement fuzzy proportional-integral-derivative n/a FG n/a Reuse
controllers to control temperature and humidity
Xu et al. (2018a) Control Apply adaptive two timestep receding horizon optimal Venlo CO2, H Lettuce Reuse
control (TTRHOC) on a greenhouse
Xu et al. (2018b) Control Realize economic optimization in CSG using TTRHOC CSG CO2, H, LED Lettuce Extension
Xu et al. (2019) Control Quantify the benefits of TTRHOC on a greenhouse with Venlo CO2, H, LED Lettuce Extension
LEDs
Zhang et al. (2020) Analysis Develop a model with dynamic cover absorbance and Venlo – None New
transmittance factors
are modelled similarly. where F (− ), called the view factor, expresses how visible the two objects
are to each other; ε1 and ε2 (− ) are the emissivities of the two objects,
2.4.4. Thermal radiation which are a property of the bodies' material; σ = 5.67 ⋅ 10− 8 W m− 2 K− 4
Thermal radiation between two objects, also called long wave radi is the Stefan-Boltzmann constant; and T1,K and T2,K (K) are temperatures
ation or far infrared (FIR) radiation, is modelled according to the Stefan- in kelvin of the two objects.
Boltzmann law: Air emits very little thermal radiation, so using indoor air tempera
(( )4 ) (
)4 ( ) ture as T1,K and the outdoor air temperature as T2,K in Eq. 9 is typically
QFIR = Fε1 ε2 σ T1,K − T2,K W m− 2 (9)
insufficient to calculate the thermal radiation losses of the greenhouse
system. The effective sky temperature (or simply sky temperature),
10
D. Katzin et al. Agricultural Systems 198 (2022) 103388
Fig. 4. Classification of studies related to process-based greenhouse climate modelling. Ornam: ornamental crops. CSG: Chinese solar greenhouse. PE: polyethylene.
HPS: high-pressure sodium lamps. LED: light-emitting diodes. MHL: metal halide lamps. n/a: information was not available.
11
D. Katzin et al. Agricultural Systems 198 (2022) 103388
Table 3
Decomposition of greenhouse models, including model objective, number of objects considered in convective and radiative exchanges, variables influencing selected
processes, and origins of model components. Color code: Blue: model extension, components that were changed during model extension based on previous sources.
Green: new model or new model component. Red: model reuse. Yellow: component not included in the model. Model objective: Ctrl: climate control, Expl:
exploratory modelling, Calib: calibration, Analys: system analysis. Greenhouse type: PE: polyethylene, CSG: Chinese solar greenhouse, Ven: Venlo. Development:
Ex: extension, RE: reuse. Model decomposition: C: calculated based on other components, F: only fogging included, G: given as input, O: constant, N: no heating
contribution, New: new model or component, ∝: proportional to input, ⨯: component excluded, ✓: component included. Lamp type: n/a: not specified, H: high-
pressure sodium, L: LED, M: metal halide. Influence of climate variables: Ci: indoor CO2 concentration, DW: crop dry weight, Hi: indoor humidity, Ho: outdoor
humidity, LAI: leaf area index, n/a: information not available, R: radiation, Tc: crop temperature, Ti: indoor temperature, To: outdoor temperature, V: ventilation
control, W: wind. [− ] model simplification. [+] model extension. References: [1] Abbes et al. (2010). [2] Ahamed et al. (2018b). [3] Jolliet et al. (1991). [4] J. Chen
et al. (2016). [5] J. Chen et al. (2015). [6] Pasgianos et al. (2003). [7] De Zwart (1996). [8] Jones et al. (1999). [9] Bontsema et al. (2007). [10] Stanghellini et al.
(2012). [11] Marcelis et al. (2009). [12] Blasco et al. (2007) [13] Vanthoor et al. (2011b). [14] Vanthoor et al. (2011a). [15] Lammari et al. (2012). [16] Albright et al.
(2001). [17] Van Ooteghem (2007). [18] Stanghellini (1987). [19] Hasni et al. (2011) [20] Seginer et al. (2018). [21] Seginer et al. (2020a). [22] Tiwari and Goyal
(1998). [23] Tap (2000). [24] Roy et al. (2002). [25] Goudriaan and Van Laar (1993). [26] Van Henten (2003).
which is a function of the thermal radiation emitted from the sky to between greenhouse objects, such as the greenhouse cover, the crop, and
wards the earth, is typically used for T2,K when calculating the radiative the soil, are modelled similarly.
losses of buildings (Evangelisti et al., 2019), including greenhouses. Sky
temperature may be given as an input, or calculated based on outdoor 2.4.5. Energy losses to latent heat
air temperature, humidity, and cloud cover. As in the case of convection, An important component of the energy balance is conversion of
the cover temperature is often used for T1,K. Thermal radiation exchange sensible to latent heat. This is described as:
12
D. Katzin et al. Agricultural Systems 198 (2022) 103388
Fig. 5. Inheritance chart of recently published greenhouse modelling studies and the models on which they are based. Studies are sorted vertically by year of
publication, with the most recent studies in the bottom and the oldest ones on top. Colored boxes indicate studies included in this review. Out of space considerations,
only first author names are given, except in cases of possible ambiguity. CSG: Chinese solar greenhouse; HPS: high-pressure sodium lamps; LED: light-emitting diodes;
PV: photovoltaic cells.
13
D. Katzin et al. Agricultural Systems 198 (2022) 103388
Table 4
Validation of greenhouse models. See Section 3.4 for the definitions used here for the seasons, validated variables, and validation metrics. Variables: Ci: indoor CO2
concentration, DW: harvested dry weight, E: electricity produced, FW: harvested fresh weight, H: heating energy, Hi: indoor humidity, R: radiation, RHi: indoor
relative humidity, Tc: cover temperature, Ti: indoor temperature, Tp: plant temperature, Ts: soil temperature, Tw: wall temperature, V: ventilation rate, Wi: indoor
wind speed. Validation metrics: d: Wilmott's index of agreement, EF: model efficiency, MAPE: mean absolute percentage error, maxE: maximum error, maxRE:
maximum relative error, ME: mean error, MRE: mean relative error, MSE: mean squared error, NSE: Nash-Sutcliffe's coefficient of efficiency, RE: relative error, RMSE:
root mean squared error, rRMSE: relative RMSE, SPC: square of the Pearson correlation coefficient, TSSE: total sum of squared error. n/a: information was not
available. All studies included plotted graphs of validation results. See Section 3.4 for definitions of commonly used metrics, and the respective publications for
definitions of metrics used in only one study, marked with (*).
Reference Location and season of Facility of evaluation Validated Dataset duration Validation metrics Main validation results
evaluation (size) variables (sampling rate)
Abbes et al. Borj Cedria, Tunisia. Research (100 m2) R, RHi, Ti 19 days (n/a) MSE, NSE R: NSE = 0.98, MSE = 39 W m− 2. RHi:
(2019) Spring MSE = 0.45, NSE = 0.98. Ti: MSE =
3.63 ◦ C, NSE = 1
Ahamed et al. Elie, Manitoba, Commercial (210 m2) H, R, Ts, Tw 3 days (10 min – 1 maxE, ME, MRE, H: rRMSE = 11.5%. R: NSE = 0.71,
(2018a) Canada. Spring h) NSE, RMSE, RMSE = 68.34 W m− 2 Ts: NSE = 0.68,
rRMSE, RMSE = 1.8 ◦ C
Ahamed et al. Saskatoon, Canada. n/a (1125 m2) H, R 8 months (month) maxRE, MRE, NSE, H: MaxRE = 9%, MRE = 4.6%. R: NSE =
(2018b, 2018c) Spring-autumn RMSE 0.78, RMSE = 112.61 W m− 2
Alinejad et al. Shiraz, Iran. Full year Commercial (4081 E, H, R, RHi, 1 year (monthly) maxRE R: maxRE <8%, Ti: maxRE <3%,
(2020) m2) Ti, V
C. Chen et al. Beijing, China. Winter n/a Ti 2 months (hourly) d (*) Ti: d = 0.987
(2019)
J. Chen et al. Hangzhou, China. n/a (230 m2) R 8 h (hourly) MRE, RMSE R: RMSE = 12.61–21.97 W m− 2
(2019) Autumn
De Ridder et al. Belgium. Autumn Research (160 m2) Ti 2 × 7 days (20 min) NSE Ti: NSE > 90%
(2020)
2
Esmaeli and Shenyang, China. n/a (756 m ) Ti 1 day (hourly) maxE, NSE, RMSE Ti: maxE = 2.8 ◦ C, NSE = 0.95, RMSE =
Roshandel Winter 0.32 ◦ C
(2020)
Golzar et al. Conthey, Switzerland. Research (360 m2) H, DW 10 months (week- NSE, PBIAS (*), DW: NSE = 0.96, PBIAS = 0.18, rRMSE
(2018) Winter-autumn month) rRMSE = 23%. H: NSE = 0.91, PBIAS = 0.18,
rRMSE = 27%
Hemming et al. Bleiswijk, The Research (96 m2) FW 3 months (weekly) – –
(2019b) Netherlands. Summer-
autumn
Jomaa et al. Borj Cedria, Tunisia. Research (100 m2) Hi, Ti 3.5 days (n/a) – –
(2019) Spring
Katzin et al. Bleiswijk, The Research (144 m2) Ci, H, RHi, Ti 112 days (5 min) ME, RE, RMSE, H: RE = − 0.92-11.6%. RHi: RMSE =
(2020b) Netherlands. Autumn- rRMSE 5.52–8.5%. Ti: ME = − 0.09-0.05 ◦ C,
winter RMSE = 1.74–2.04 ◦ C
Lammari et al. Avignon, France. Research (n/a) Hi, Ti 2 × 7 days (n/a) – –
(2012, 2020) Spring-summer
Ma et al. (2019) Indiana, USA. Autumn Research (n/a) R, Ti 7 days (minute) NSE R: NSE = 0.9, Ti: NSE = 0.88
Mohammadi et al. Tabriz, Iran. Autumn Research (15 m2) Tc, Ti, Tp, Ts 8 h (minute) EF (*), MAPE (*), Tc: SPC = 0.96, RMSE = 2.21 ◦ C. Ti:
(2020) RMSE, SPC, TSSE SPC = 0.98, RMSE = 1.64 ◦ C. Ts: SPC =
(*) 0.98, RMSE = 1.84 ◦ C
Pérez-González Guadalajara, Mexico. Research (30 m2) RHi, Ti 1–3 days (second) J (*) J = 5.64–9.4327
et al. (2018) Autumn, spring
2
Rasheed et al. Daegu, South Korea. Research (168 m ) Ti 20 days (n/a) NSE Ti: NSE = 0.79–0.84
(2019) Winter, autumn.
Righini et al. Klepp, Orre, Norway. Commercial (5760 FW, Ti FW: 3–7 months; rRMSE, MRE FW: MRE = 0.7–4.3%. Ti: rRMSE =
(2020) Winter, spring, m2), experimental (n/ Ti: 3×(6–8) days 7.1–9.6%
summer a) (n/a)
Sethi (2019) Ludhiana, India. Research (100 m2) R, Tc, Ti, Tp 7–12 h (hourly) RMSE R: RMSE = 4.01 W m− 2. Tc: RMSE =
Summer, winter. 3.91 ◦ C. Ti: RMSE = 4.69 ◦ C. Tp: RMSE
= 3.7 ◦ C
Su et al. (2018) Chongming, Shanghai, n/a (875 m2) Ci, Hi, Ti 5 months (5 min) RMSE Ci: RMSE = 50 mg m− 3 Hi: RMSE =
China. Autumn-spring. 1.68 g m− 3. Ti: RMSE = 2.1 ◦ C
Zhang et al. Taian, Shandong, Prototype (15 m2) Tc, Ti, Ts 3 × 15 days (5 min) ME, RMSE, SPC Ti: SPC = 0.98, RMSE = 1.36–2.01 ◦ C.
(2020) China. Winter, spring. Ts: SPC = 0.66–0.99, RMSE =
0.1–1.93 ◦ C
( )
QLatent = L⋅WLatent W m− 2 (10) ventilation WVent (Eq. 6) is not in itself associated with a change in the
energy balance.
where L (J kg− 1) is the latent heat of evaporation of water, and:
( ) 2.4.6. Transpiration
WLatent = WTrans + WEvap − WCond kg m− 2 s− 1 (11) A wide range of approaches can be used for modelling crop tran
spiration, as outlined in detail by Katsoulas and Stanghellini (2019).
is the net amount of water transformed to vapor in the system: WTrans is
These approaches range from an empirically fitted function where
water transpired by the crop, WEvap is water evaporated from the soil,
transpiration depends only on solar radiation, through aerodynamic
and WCond is vapor condensed to water on cold surfaces such as the cover
models that include the influence of wind, to detailed models that
or screens. For each of these components W, the associated energy flow
include the energy balance of the crop and the response of stomata to
is L ⋅ W. Not every change in the water vapor balance is associated with
environmental attributes including total radiation intercepted by the
latent heat exchanges: for example, loss of water vapor through
14
D. Katzin et al. Agricultural Systems 198 (2022) 103388
crop, vapor pressure deficit (VPD), air temperature, and CO2 is colder than the dew point of the air. Equivalently, this means that the
concentration. saturation vapor pressure at the temperature of the surface is higher
One of the simplest ways to model transpiration is to treat it as a than the vapor pressure of the air. Condensation typically occurs on the
linear function of radiation: indoor side of the greenhouse cover or on screens, but may also happen
( ) on the crop itself, the floor or soil. An equation to describe condensation
WTrans = A0 ISun + B0 kg m− 2 s− 1 (12)
is:
Where ISun (W m− 2) is radiation from the sun, and A0 (kg J− 1) and B0 { ( )} ( )
WCond = max 0, g VAir − VSurface kg m− 2 s− 1 (16)
(kg m− 2 s− 1) are fitted parameters that may depend on the crop, the crop
stage, or the growing season. where VAir (kg m− 3) is the vapor concentration of the air, VSurface (kg
Another common way to model transpiration is the Penman- m− 3) is the saturation vapor concentration at the temperature of the
Monteith formula (Monteith, 1965), which describes the latent heat of surface, and g (m3 m− 2 s− 1) is an exchange coefficient. This coefficient g
evaporation from a leaf (shown here using the notation of Katsoulas and may be related to the heat exchange coefficient h in Eq. 8, for instance it
Stanghellini (2019)): may be proportional to it.
ΔRn + ρCp Di ga ( )
QTrans,leaf = W m− 2 {leaf} (13) 2.4.8. Crop photosynthesis
Δ + γ(1 + ga /gc )
Leaf and canopy photosynthesis modelling is a broad and long-
where Δ (kPa K− 1) is the slope of relationship between saturation vapor standing discipline with an extensive range of approaches (Hikosaka
pressure and temperature, Rn (W m− 2 {leaf}) is the net radiation inter et al., 2016). Earlier reviews described some of the crop modelling ap
cepted by the crop, ρ (kg m− 3) is the density of air, Cp (J kg− 1 K− 1) is the proaches used in horticulture (Gary et al., 1998) and in greenhouses in
specific heat of air, Di (kPa) is the vapor pressure deficit of the air, ga (m particular (Marcelis et al., 1998).
s− 1) is aerodynamic conductance, γ (kPa K− 1) is the psychrometric As with transpiration, photosynthesis models range from very simple
constant, and gc (m s− 1) is stomatal conductance. Transpiration models where only light is taken into account, to complex models
WTrans,leaf in kg m− 2 s− 1 is then calculated by dividing QTrans,leaf by the including the influence of temperature, CO2 concentration, and crop
specific heat of water evaporation, see Eq. 10. processes such as assimilate demands of the various organs. A yield
Stanghellini (1987) modified the Penman-Monteith formula for the model may also be included, predicting how much produce can be sold
case of greenhouse crops. First, a factor for crop leaf area index (LAI, m2 by the greenhouse, and when. Again, the level of detail varies consid
{leaf} m− 2 {floor}), which expresses the leaf area of the crop per area of erably between such models (Kuijpers et al., 2019). The type of crop may
greenhouse floor, was used in this case to convert leaf transpiration to influence the level of detail: leafy crops such as lettuce typically require
crop transpiration: less detail than fruiting crops such as tomato or cucumber.
As with transpiration, a description of the leaf area index (LAI) and
ΔRn + 2⋅LAI⋅ρCp Di ga ( )
QTrans = W m− 2 {floor} (14) its development through time may also be included in the crop model. In
Δ + γ(1 + ga /gc )
any case, LAI is typically an important component of the photosynthesis
Furthermore, calculation of the aerodynamic conductance ga was model. This means that even if a simple photosynthesis model is used,
modified to describe a greenhosue environment. Lastly, stomatal some assumption or estimate regarding LAI is needed.
conductance gc was calculated following the approach of Jarvis (1976),
where stomatal conductance depends on environmental factors such as
2.5. Time scales in greenhouse climate models
radiation, vapor pressure deficit, temperature, and CO2 concentration:
( )
gc = gM f1 (Rn )f2 (Di )f3 (Ti )f4 (CO2 ) m s− 1 (15) The dynamics of the greenhouse climate are influenced by processes
which may be classified as “fast” and “slow” (e.g., Tap et al., 1993; Van
The notation here follows, again, Katsoulas and Stanghellini (2019) Henten, 1994; Van Straten et al., 2010). The slow variables are typically
(see Eq. 13). Here, gM (m s− 1) is the maximal stomatal conductance, and related to crop processes, which evolve in the order of days, weeks, and
f1, f2, f3, f4 are unitless functions with values between 0 and 1 that months. The fast variables are related to climate variables, which can
represent the influence of, respectively, radiation, vapor pressure evolve in the order of seconds, minutes, and hours. These widely
deficit, temperature, and CO2 concentration on stomatal conductance. different time scales often cause inaccurate or inefficient results during
Villarreal-Guerrero et al. (2012) discussed the differences between the numerical integration (i.e., during greenhouse simulation), especially in
Penman-Monteith and the Stanghellini transpiration models, and studies related to optimal control (Tap et al., 1993; Van Straten et al.,
compared their performances against measured data. 2010).
In the equations above, Ti may refer to either the air or the crop Thus, a common approach is to separate the state variables of a
temperature, and Di may refer to either the vapor pressure difference of greenhouse into fast and slow time scales, and compute these two time
the air or the vapor pressure difference between the saturated vapor scales separately. The fast variables may be assumed to “achieve a
pressure at the temperature of the crop and the saturated vapor pressure steady-state infinitely fast” (Van Henten and Bontsema, 2009). In effect,
of the air. If crop temperature is used in the calculation of transpiration, this results in setting the differential equations governing the fast states
then naturally this attribute must also be included in the greenhouse to be constantly zero. This decomposition into fast and slow dynamics
model. Similarly, if LAI is included in the calculation of transpiration, helps circumvent many of the numerical issues during integration, and
some estimate, assumption, or model describing LAI must also be results in more efficient computation. Furthermore, this decomposition
included. helps analyze and understand the behavior of the greenhouse system
Another approach for modelling transpiration is based on the as along the different time scales (Van Henten and Bontsema, 2009).
sumptions that the indoor water vapor concentration is in steady state, Nevertheless, a distinction should be made between how a model
and that all vapor flows besides ventilation are negligible. Considering describes the dynamics of a greenhouse system, and how this description
Eq. 1, this leads to WTrans = WVent, which allows to estimate transpiration is used when the system is solved or integrated in practice. For example,
based on ventilation rate and indoor and outdoor vapor concentrations the Van Henten model (Eq. 2) may be used in some cases without any
(Eq. 6). time scale decomposition (Van Henten, 2003), while in other cases such
a decomposition proves useful (Van Henten and Bontsema, 2009).
2.4.7. Condensation Similarly, during the development of the De Zwart model (Eq. 4), a
Condensation occurs when humid air is in contact with a surface that choice was made to compute some variables using “a static equation”,
15
D. Katzin et al. Agricultural Systems 198 (2022) 103388
since “the dynamics [...] [are] of a small time base” (De Zwart, 1996). In extension: models explicitly using a single previous model and adding
other words, it is assumed that some of the climate dynamics are faster components to it; and translation: previously published models pre
than others, namely, those related to the air above the screen and to sented using new code or a new software platform.
other objects with a small heat capacity. Again, the assumption is that
these faster states achieve a steady state infinitely fast, which results in 3.2. Composition of process-based greenhouse climate models
an equation with a left-hand side equal zero. This includes for example
the temperature of the screen and the temperature of the air above the In order to get a better understanding of the differences between
screen, represented by the second and third lines of Eq. 4. However, models that are currently used, the models were analyzed and compared
since the dynamics of these variables are still described, one could with regards to how they handle the various greenhouse model com
choose to use the De Zwart model and reject this “static” assumption, if ponents. These components included:
the need arises.
Alternatively, some models choose to completely avoid describing • Energy balance: heating from the sun, the heating system, and
the slow states. For example, Van Beveren et al. (2015a, 2015b) assumed supplemental lighting; number of objects considered in thermal ra
that the crop, and specifically the LAI, remained constant throughout the diation exchanges; number of objects considered in convective and
simulated season. Conversely, Vanthoor et al. (2011c) combined a crop conductive exchanges; ventilation; latent heat.
and greenhouse model, where despite these two models operating in • Water vapor balance: transpiration (variables influencing it and
different time scales and both containing numerous state variables, all model used), evaporation and condensation (number of objects
states were directly and simultaneously calculated using differential considered).
equations. • CO2 balance: photosynthesis and yield (factors influencing them,
and model used).
3. Review of recent greenhouse modelling studies: Methodology • Leaf area index (LAI).
In order to examine the current state of greenhouse modelling, a For each component, we noted whether the model introduced a new
literature search was performed on Clarivate's Web of Science (www. method for calculating the component, or if a previous study was used.
webofscience.com). Since the term “greenhouse” is widely used in We noted which previous work was used for a specific component, if this
non-horticultural contexts, a search for simply “greenhouse model” or was noted. If a component based on previous works was modified
“greenhouse AND model” was unfeasible, yielding over 8000 results, (simplified or extended), this was also noted.
most of them irrelevant. Instead, the search term chosen was: (“green
house model*” OR “greenhouse simulation*” OR (greenhouse* AND 3.3. Inheritance of process-based greenhouse climate models
(“yield model*” OR “thermal model*” OR “heating model*” OR “yield
simulation*” OR “thermal simulation*” OR “heating simulation*” OR In order to explain the differences between current models, the
“optimal control”)) NOT (“greenhouse gas*” OR “greenhouse emis models were further investigated by examining which previous models
sion*” OR “greenhouse effect*”)). The search was performed on October they were based on, and how they had modified or combined them. This
6, 2020, for articles in the Web of Science Core Collection, published in was done based on the authors' descriptions, together with our own
2018–2020, whose topic (title, abstract, keywords, and Keywords Plus) comparison of the published model with previously published models.
matched the search term. This three-year period was chosen as repre The models included in this study were compared against their reported
sentative of the current state of the art in greenhouse modelling. The “parent” to see whether components were added or modified, and how.
search yielded 80 results. Of these, 48 articles were excluded: 15 that The parent models were also checked to see whether they themselves
described models of components of the greenhouse system; 12 that were based on common earlier works. We also noted whether some of
discussed unrelated topics; 7 that discussed greenhouse dryers; 5 that the same authors were involved in the publications describing the parent
described real world, scaled down models of greenhouses; 5 review and the daughter models. For this purpose, promotors and supervisors
studies; 3 descriptive models; and 1 CFD model. Thus, 32 articles pub were considered as coauthors of PhD dissertations.
lished in 2018–2020 were considered in this study.
3.4. Validation of greenhouse models
3.1. Overview of greenhouse modelling studies
For all models that presented a validation simulation, the following
The studies in the 32 articles were analyzed as follows: first, the attributes were collected:
study objectives were divided into 5 categories (see Section 2.2),
including exploratory modelling, model-based control, model-assisted • Location and season where measured data was collected. Here,
design, and systems analysis. Another category was included for meteorological seasons of northern latitudes were used: Winter:
studies focusing on model calibration methods. Next, a more specific December–February; Spring: March–May; Summer: June–August;
purpose of each study was summarized according to the authors' de Autumn: September–November (all studies considered were per
scriptions. The greenhouse type, crop, and equipment modelled were formed in the northern hemisphere).
described based on the authors' descriptions or, when those weren't • Type and size of the facility where data was collected, if it was
provided, on the conditions used during model validation. For green provided: research greenhouse, commercial greenhouse, or scaled-
house type, we defined a Venlo greenhouse as an even-span greenhouse down prototype.
with glass cover and walls. A Chinese solar greenhouse (CSG) was • Validated variables, i.e., variables that were predicted by the model
defined as a low greenhouse with a northern wall which provides and compared against measured data.
diurnal heat storage. A southern cover arching from the northern wall to • Duration of time in which data was collected, and sampling rate of
the southern edge of the CSG is equipped with a thermal blanket or data points.
screen that maintains heat in the greenhouse overnight. • Metrics used for validation.
Lastly, studies were categorized according to the development status • Main validation results, rounded to 2 decimal points.
of the models used: new: newly presented models (possibly combining
several previous works, but not explicitly derived from a single previous Metrics used for validation varied considerably between studies, as
model); reuse: exact copies of a previous model; parametrization: well as the terms used to describe these metrics. The definitions below,
models combining a single previous model with a new set of parameters; following Legates and McCabe (1999), were used when noting the
16
D. Katzin et al. Agricultural Systems 198 (2022) 103388
validation metrics. Here, we denote by yi the measured values, fi the 4. Results: Current state of the art in greenhouse modelling
predicted values, y the mean of the measured values, f the mean of the
∑
predicted values, summation over i, and max the maximum over i, 4.1. Overview of greenhouse climate modelling studies
where i = 1, …, n is an index ranging over measurements and corre
sponding predictions. The studies considered in this review vary in their purposes, types of
greenhouse considered, and equipment included in the greenhouse. The
∑( )
majority of the studies considered (11 of 32) focus on greenhouse
• Mean error: ME = 1n fi − yi
• Maximum error: maxE = max {fi − yi} climate control (Fig. 4). Studies focused on design are also common (8 of
• Mean relative error: MRE = y⋅n 1
∑(
fi − yi
) 32), followed by systems analysis (6), exploratory modelling (4), and
{ } model calibration (3), although the distinctions between these cate
• Maximum relative error: maxRE = max fi −y yi gories was not always clear (Table 2).
∑( )2 In studies focused on control, various methods for climate control
• Mean squared error: MSE = 1n f i − yi were described and evaluated, such as adaptive control, fuzzy logic, and
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
)2̅
• Root mean squared error: RMSE = 1n
∑(
fi − yi more (Table 2). Model-based design was used for designing a complete
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
)2̅ greenhouse system or specific components such as placement of PV cells.
∑(
• Relative root mean squared error: rRMSE = 1y 1n fi − yi Systems analysis was used to analyze the greenhouse energy use, envi
• Square of the Pearson correlation coefficient: SPC = ronmental impact, or the model itself. Exploratory modelling was used
∑
( (yi − y)(fi − f ) )
2 to test scenarios with different designs, control strategies, and equip
∑
(yi − y)
2∑
(fi − f )
2 ment. Model calibration studies proposed new methods for calibration
∑
(yi − fi )
2 or identification, such as particle swarm optimization and deep
• Nash-Sutcliffe's coefficient of efficiency: NSE = 1 − ∑ 2 networks.
(yi − y)
More than a third of the studies (11 out of 32) presented new models,
Some ambiguity was found with respect to the terms “coefficient of and the same number of studies was devoted to extensions of previous
determination”, “correlation coefficient”, “explained variation”, models (Fig. 4). Reuse of existing models was less common (9 out of 32
“explained variability”, “r2”, and “R2”. These terms were often used to cases). One study presented a translation of an existing model to a new
mean either the SPC or the NSE as defined above. In some cases, no software platform. New models were created to describe new types of
reference or equation was given, which made it difficult to determine greenhouses (e.g. CSGs, asymmetric overlap roof), to implement models
which of the metrics was actually used. In these cases, we assumed that on specific platforms (e.g. TRNSYS), to include new technologies (lamps,
the NSE was used. Whenever equations were given for the metrics used, heat harvesting), or to incorporate detailed model components (cover
they were reported according to the definitions above. For other metrics, absorbance, thermal screens) (Table 2). Notably, 6 studies defined the
if no definition was given, the term given by the authors was used. development of the model itself as a study purpose, and 8 studies defined
Some words are in order regarding terminology. It has been noted the use of a certain methodology as a purpose.
that the term “validation” may be interpreted as aiming to provide a In the majority of cases (11 of 32), the crop in the modelled green
clear yes or no answer regarding the adequacy of a model: the model is house was tomato (Fig. 4). This could be expected, as tomato is by far the
either valid or it is not. For this reason, some have suggested to avoid the most widely produced and exported vegetable in the world, excluding
term “validation” and use “evaluation” instead (Wallach et al., 2019a). potato and melons (Rabobank, 2018). The low representation (3) of
Nevertheless, the two terms are often used interchangeably. In this ornamental crops is surprising, as they make up at least half of the
study, we use the term “validation” in a narrow sense, as the comparison world's greenhouse production (Stanghellini et al., 2019). Remarkably,
of model predictions against measured data. The term “evaluation” is several models (5) were designed or evaluated under the assumption
understood in a broader sense, which includes validation but may also that the greenhouse does not house any crop.
include other components, such as the process of data collection. The types of modelled greenhouses could broadly be classified into
three: Venlo glasshouses, Chinese solar greenhouses, and polyethylene
3.5. Studies used in each analysis greenhouses (Fig. 4). CSGs varied in the type of covering material used,
and polyethylene greenhouses varied in shape (Table 2). The modelled
In the investigations of model composition and inheritance, models greenhouses also varied considerably in the equipment they included,
based on software platforms such as TRNSYS, EnergyPlus or COMSOL with heating the most common equipment included (19 of 32 studies),
(see Section 2.1) were excluded from the analysis, since their de followed by lighting (15 studies in total), CO2 injection (12), fogging (7),
scriptions were typically insufficient for users unfamiliar with these and thermal screens (7) (Fig. 4).
programs to understand the inner workings of the model. In the analysis
of model inheritance, only complete greenhouse models were consid 4.2. Composition of process-based greenhouse climate models
ered as eligible parent models. For the sake of clarity, the inclusion or
combination of model components, such as a separate transpiration or As described in Section 2.4, greenhouse climate models describe a
convection model was not described in the inheritance chart. Naturally, large number of processes, and a vast range of approaches is available
for the analysis of model validation, only studies presenting any kind of for each process. Among the models analyzed in the current studies,
validation were considered. these differed considerably in how they incorporated the various
For some studies included in this review, the details of the model greenhouse system components (Table 3). At the same time, there is
described were given in a previous publication. In these cases, we used considerable overlap between the models analyzed, which could
information from the previous publication to analyze the model possibly result in redundancy. A considerable amount of model devel
composition, inheritance and validation. This was done for Ahamed opment (14 of 24 models analyzed) consisted of putting together pre
et al. (2018c) by also considering Ahamed et al. (2018b); for Lammari viously published components; less than half of the studies (10 of 24)
et al. (2020) by including Lammari et al. (2012); and for Esmaeli and described newly developed model components. The solar heat load was
Roshandel (2020) with Esmaeli and Roshandel (2017). typically given as a model input or set as a fixed proportion of an input.
Heating was typically given as an input or calculated based on other
energy fluxes. When lighting was included, the resulting heating load
was often assumed to be proportional to the lamp power input, no
matter the lamp type (6 out of 9 cases). In 2 cases describing LEDs, it was
17
D. Katzin et al. Agricultural Systems 198 (2022) 103388
assumed that light does not contribute any heat to the greenhouse. The considered: the facilities used for data collection, the validated vari
number of objects included in thermal radiation (FIR) exchanges varied ables, the dataset length and sampling rate, and in validation metrics
from 0 to 11, and the number of objects included in convective ex (Table 4). The size of the facility used varied from a 15 m2 prototype to a
changes varied from 2 to 18. Ventilation was neglected (2 out of 24 5760 m2 commercial greenhouse. The validated variables typically
cases), assumed constant (5 cases), given as an input (8 cases), or included the indoor temperature, but other variables were also consid
calculated based on various variables (9 cases). Latent heat was not ered, such as indoor humidity, radiation inside the greenhouse, and
included in 5 out of 24 models, and in 3 only the influence of fogging was heating energy used. Dataset length varied from less than a day to a full
included. LAI was excluded (9 of 24 cases) or assumed constant (8 cases) year, and the sampling rate of measurements varied from one second to
by a majority of the models. Approaches to modelling transpiration one month. Overall, the sizes of datasets used for model validation were
varied considerably, as previously noted by Katsoulas and Stanghellini rather small, with datasets representing the major part of a year (8–12
(2019). Relatively few models considered photosynthesis (9 of 24) and months) only consisting of weekly or monthly measurements. In terms of
yield (8 of 24), and approaches varied between those that did. validation metrics, no common standard was found. All studies included
Regarding the time scales included in the models, all models that did graphs comparing measured and simulated values. Some studies added
not include a yield component (16 of 24) considered only the fast dy no information beyond the presented graphs, while others used multiple
namics of the greenhouse climate and neglected the slow dynamics of metrics to analyze the model predictions. RMSE was the most common
the crop. In 6 out of the remaining 8 models, both the fast dynamics of metric for validation (used in 9 of the 22 studies), followed by NSE (8
the greenhouse climate as well as the slow dynamics of crop develop studies). Other metrics were also used, with some used only by a single
ment were included. In the model of Seginer et al. (2020b), the crop was study. Because of this wide range of evaluation methods, differing in
assumed to be mature, and therefore “steady”, with yield directly pro timespan, sampling rate, validated variables and validation metrics, it is
portional to net photosynthesis. Similarly, Katzin et al. (2020b) assumed practically impossible to compare model performance based solely on
that the crop was mature, and therefore LAI and crop development validation results.
status were constant.
5. Discussion
4.3. Inheritance of process-based greenhouse climate models
5.1. What is the source of variation between greenhouse climate models?
Some of the differences and similarities between models can be
explained by the model inheritance chart (Fig. 5). For example, it can be The vast range of crops, structures, equipment and climates that
seen that the models of Katzin et al. (2020b) and of Righini et al. (2020) characterizes greenhouse horticulture does not seem sufficient to
were both extensions of Vanthoor et al. (2011b) and that the models explain the variation between greenhouse models. The majority of
described in Pérez-González et al. (2018) and in Lammari et al. (2020) studies reviewed considered a tomato crop, while many others assumed
both originated with Boulard et al. (1996). The models used by an undefined generic crop or no crop at all (Fig. 4). Adding new
Mohamed and Hameed (2018), Gharghory (2020), and Subin et al. equipment to an existing greenhouse model is sometimes a motivation
(2020) all originated with the model of Albright et al. (2001). It can also for model extension (Fig. 5), but that in itself does not seem to justify the
be seen that several early models (Albright et al., 2001; Boulard et al., development of entirely new models. Regarding structure type, most
1996; De Zwart, 1996; Vadiee and Martin, 2013; Van Henten, 1994; Van studies were concerned with Venlo type greenhouses (Fig. 4), although a
Ooteghem, 2007) are still used as a basis for many recent studies. These considerable number of studies focused on CSGs, and several were
models are based on even earlier models (Bot, 1983; Boulard and Baille, devoted to transforming an existing model to a CSG (Fig. 5). This seems
1987; Garzoli, 1985). Some models combine several earlier models (e.g., to indicate a growing interest in the modelling of CSGs. It remains to be
Lammari et al., 2012 combines Albright et al., 1985; Boulard et al., seen whether this trend will continue. Arguably, this will depend on
1996; Draoui, 1994), while others are reportedly based only on a single developments in the global greenhouse sector, and the prevalence of
previous model (e.g., Tchamitchian, 1992 is based on Udink ten Cate, CSGs compared to other types of greenhouses.
1983, and Taki et al., 2016 is based on Van Ooteghem, 2007). The model Modelling objective provides a better, although partial, explanation
of Tiwari, 2003 was used to convert the model of Ahamed et al., 2018b to model variation: the most complex models found in this study, with
to a CSG (Ahamed et al., 2018a). At the same time, for several recent 16 or more objects included in convective exchanges, were all devoted
studies (Ahamed et al., 2018a; C. Chen et al., 2019; De Ridder et al., to exploratory modelling (Table 3). At the other end of the complexity
2020; Sethi, 2019) we were not able to identify whether they were based scale, studies focused on calibration and control tended to use simple
on previous greenhouse models, and if so, on which. models, with few objects included in the convective and FIR exchanges.
In many cases where a model was reused or extended, authors of the Out of 11 models that were found to have 3 or less convective objects, 7
parent model were also involved in the new study. In 7 out of 11 cases of dealt with control and 2 with calibration (Table 3). Studies on calibra
model extension, the original and extended model shared coauthors. The tion and control tended to focus less on accurate predictions, and often
same was true for 4 out of 9 cases of model reuse. Cases of model reuse or lacked validation (Table 4). The approach in these studies was also more
extension by authors unrelated to the original publication are limited to generic compared to exploratory studies, often providing little detail on
quite simple models, as can be seen by their decomposition in Table 3: the type of structure or crop, and even neglecting the crop completely
Xu et al. (2018a) reused the model of Van Henten (2003); Mohamed and (Table 2).
Hameed (2018) reused the model of Albright et al. (2001); Gharghory Thus it may be argued that models devoted to exploratory modelling
(2020) and Subin et al. (2020) reused the model of Pasgianos et al. require a high level of complexity while models devoted to control and
(2003); and Jomaa et al. (2019) reused the model of Blasco et al. (2007), calibration require a low level of complexity. If this were the case, it
while using parameters from a previous study. Model extension without would explain some of the variation in greenhouse modelling. However,
shared coauthors is also reserved for relatively simple models, e.g. the exceptions do exist, with the relatively simple model of Ahamed et al.
extensions by Xu et al. (2018b, 2019) to Van Henten (2003), the (2018a), with only 5 convective objects, devoted to exploratory
extension by Su et al. (2018) to Tap (2000), and the extension of Pérez- modelling, while the relatively complex model of Abbes et al. (2019),
González et al. (2018) to Hasni et al. (2011). with 8 convective objects, dedicated to control. Of course, one may also
argue that the number of convective objects is not the only way to
4.4. Validation of greenhouse climate models quantify model complexity, a concept which remains open for
interpretation.
Models differed in the techniques used for validation in all aspects In any case, the distinction between detailed and complex
18
D. Katzin et al. Agricultural Systems 198 (2022) 103388
“prediction-focused” models and simpler “control-focused” models goes 1996; Sinclair and Seligman, 2000). Monteith (1996) advocated finding
back at least to the 1980's, when Bot (1983) and Udink ten Cate (1983) a “balance” between simplicity and complexity. He stressed that com
developed in parallel models that represented these two objectives. Tap plex models are more difficult to understand and that it is often easier to
(2000) reported that in 1986, Vennegoor op Nijhuis compared these two derive insights from simpler models, even if those may provide less ac
models as part of an MSc thesis and found them to be similar in terms of curate outputs.
predictions, with the model of Udink ten Cate being considerably faster Unfortunately, besides this general advice and the view that simple
in terms of computation time. Indeed, a preference for simpler models in models are often preferable, it is hard to find very practical advice
control applications due to computational issues remains relevant to this regarding model complexity. Passioura (1996) maintained that crop
day: see for example the discussion in Van Henten and Bontsema (2009), models should be “as simple as possible” and should have “a small
or the several technical steps required in Xu et al. (2018a) to obtain an appetite for data”. Keating (2020) reiterated a statement attributed to
accurate and fast solution to an optimal control problem. Albert Einstein that “the model should be as simple as possible in the
At the same time, it seems reasonable that models designed for context of intended application, but no simpler”. In the context of
predicting scenarios would include many processes. First of all, because greenhouse modelling, Vanthoor (2011) proposed that developing
new insights regarding particular system components and processes can simple models is often easier said than done: “According to Johan Cruijff
only be uncovered if those components and processes are included in the simple soccer is the most difficult to play, and unfortunately, this also
model. Second, there seems to be an agreement between model de applies to simple modelling”. Efforts to methodologically address the
velopers and users that model structure, i.e., the assumptions and re question of model selection, as has been done for example by Crout et al.
lationships underlying the model, are important factors when assessing (2009) in the context of environmental modelling, are rare in crop
the reliability of a model for exploratory simulations (Eker et al., 2018). modelling, and even more so in greenhouse modelling.
This view could explain why models developed for prediction tend to At the same time, there is quite some evidence to attest to the power
incorporate many process-based relations based on physical and bio of simple models: Stockle (1992) showed that a photosynthesis model
logical principles, rather than summarize phenomena with simpler can be considerably simplified, and thus be made easier to use and un
functions: the inclusion of phenomena which are understandable and derstand, without meaningfully increasing its prediction error. Soltani
trusted by a potential user (or the model developer) promotes trust in and Sinclair (2015) demonstrated that simple crop models sometimes
the model and its predictions. provide more accurate predictions than complex ones. In the context of
Another factor that could contribute to model variation, and may be greenhouse crops, the case of the TOMGRO tomato crop model is
related to the modelling objective, is the issue of time scales. As remarkable, where the number of state variables was reduced from 574
mentioned in Section 2.5, including different time scales in a model can to 5, while still achieving good predictions (Jones et al., 1999). At the
result in computational difficulties, especially in studies devoted to same time, the success of such model reduction endeavors strongly de
optimal control. Thus, control-focused studies may prefer to neglect the pends on the purpose of the modelling study and the sensitivity of the
slow processes of crop development and focus only on the fast climate model and its sub-processes towards this particular purpose.
dynamics of the greenhouse. Alternatively, design-focused studies may Considering the large number of processes and objects that may be
neglect some of the fast climate responses which do not have a strong present in a model (Table 1, Table 3), and the general preference for
influence on the performance of a greenhouse over a full year. simpler models, a natural question to ask is which of them should be
Nevertheless, the majority of models found in this survey, including included for any particular modelling purpose. Naturally, a modelling
models from all categories of objectives, did not include the slow crop study that examines the influence of a particular process or object (e.g.,
dynamics (Table 3). Models that did include the slow time scales were reduced crop transpiration, or the use of a thermal screen), must include
used in studies focused on exploratory modelling, greenhouse design, the process or object in question in the model. Besides that, one way to
and climate control. The four design studies analyzed all included test which other processes should be included could be by starting out
relatively few objects in the heat exchange balances, which indicates with a complex model, and testing each process individually for its in
that indeed, these studies tend to simplify the fast climate dynamics. At fluence on the intended objective. Processes that are found to have little
the same time, one design study included the slow crop dynamics, influence on the objective could then be consecutively omitted. How
showing that opinions vary regarding which time scales are the most ever, it should be noted that the influence of each process depends not
appropriate for each type of study. We see then that the issue of time only on the modelling objective, but also on the specific settings such as
scales, and the way those are treated in order to serve the modelling location of the greenhouse, type of structure, type of crop and crop stage,
objective, seems to contribute to some of the variation we see among and more. These factors will also influence which processes are domi
models. nant and which are negligible for a given modelling objective.
In any case, although modelling objective seems to give the best Another, and hopefully less cumbersome approach, would be to
explanation for the wide range in model complexity, it should be compare existing simple and complex models and test how they serve a
stressed that it is often difficult to accurately extract a given model's given modelling objective. Unfortunately, it is very difficult to compare
objective based on the authors' descriptions. In particular, the distinc the performance of greenhouse models (Section 4.4). Constructing a
tions between exploratory modelling, design, and system analysis are framework where greenhouse climate models could be compared and
not always clear-cut (Table 2). Furthermore, in some cases, the evaluated using a common dataset of measurements would help eluci
impression could arise that models were developed solely for the pur date how model complexity influences model performance. A common
pose of developing them. In other cases, the main objective seems to be framework for decomposing and comparing tomato crop models has
the demonstration of a certain methodology, and the accurate repre been presented by Kuijpers et al. (2019). It could be valuable, although
sentation or application for greenhouses seems secondary. The objec considerably more complex, to construct such a framework for green
tives assigned to the models throughout this study (Fig. 4, Table 2, house climate models.
Table 3) should be viewed with this observation in mind.
5.3. Why are so many different models being used and developed?
5.2. How complex should process-based greenhouse climate models be?
As mentioned in Section 1, the abundance of greenhouse models is
Seeing the vast range in model complexity, a question that arises is not new. More than 30 years ago, Van Bavel et al. (1985) called to focus
how complex greenhouse models should really be. This question is efforts on the improvement of existing greenhouse models, rather than
extensively debated in the context of crop modelling (Antle et al., 2014; “the writing of entirely new programs and the construction of new
Hammer et al., 2019; Keating, 2020; Monteith, 1996; Passioura, 1973, models”. Not much later, Lacroix and Zanghi (1990) urged that “rather
19
D. Katzin et al. Agricultural Systems 198 (2022) 103388
than build new models, it should usually be sufficient to take up existing 5.4. Model validation
models, adapt them to specific needs and improve them further”
(translated). These calls seem to have remained unheeded, with many 5.4.1. To what extent are greenhouse models valid?
models still being developed, several of which serving similar purposes An essential component of model development is the validation of
(Table 2). model predictions against measured data. In greenhouse modelling,
A possible explanation for the abundance of greenhouse climate evaluations vary considerably in their approaches. Details on if and how
models is the large variety in types of greenhouses. Indeed, greenhouses data was used for model development (e.g., for parameters calibration)
are used in widely different climatic environments, with various types of are often missing. In this review, the size of the measured dataset varied
flooring and soil, numerous shapes and designs, diverse covering ma considerably in the duration and sampling rate of the measured time
terials, and various cropping techniques. Nevertheless, this wide varia span, and in the majority of cases (13 of 22) data was collected in
tion in itself does not seem to justify the proliferation in models. For relatively small research facilities (Table 4). Regarding sampling rate, it
example, the abundance of models representing Dutch greenhouses and may be expected that processes such as crop yield or energy use may be
climate conditions seems incongruent with the relative homogeneity of aggregated over a longer timespan than fast processes such as the indoor
this country's climate and greenhouse sector. Furthermore, some of the climate. However, in general there was no agreement between the
varying attributes listed above could be treated by a single model by relevant time scale of a variable and the sampling rate used for its
modifying its parameters. This was done for instance by Vanthoor et al. validation (Table 4). Naturally, model developers can only work with
(2011b), who used a single model with different parametrization to the data they have. However, reflections are scarce on whether the data
represent an arch shaped multi-tunnel polyethylene greenhouse in Italy, used for evaluation is truly representative of the system being modelled
a Venlo type greenhouse in the Netherlands, a Venlo type greenhouse in (Wallach et al., 2019a).
Texas, and an arch shaped single-tunnel polycarbonate greenhouse in Considerable differences also exist in the metrics used in model
Arizona, all with satisfying results. validation (Table 4). This abundance of metrics may be seen as a positive
A more plausible reason for the current model abundance is that development, since validation studies of greenhouse models in the
some models seem to be developed simply for the sake of developing a 1980's and 1990's often provided only graphs of measured vs simulated
model (Table 2). Granted, a study may have several purposes, some values and some qualitative remarks about the model predictions such
stated more explicitly than others, and the definition of a study's purpose as “good”, “fair”, or “reasonable”. As noted earlier (Section 3.4), the
as described in Table 2 is based on subjective interpretation. Never terms currently used for validation metrics are inconsistent, with several
theless, clear motivation behind a particular model's development is different names given to the same evaluation metric, and worse – the
often missing, and it seems that model development is in itself consid same names used for two different metrics, namely, terms such as “co
ered a worthwhile research objective. This issue has been previously efficient of determination”, “correlation coefficient”, and “R2” used for
identified in the field of crop modelling (Sinclair and Seligman, 2000). both the SPC and the NSE. This confusion seems to stem from the fact
Moreover, the lack of an explicit statement of modelling objectives that when a model is derived using a linear least squares regression, the
makes it difficult to judge whether a given model serves its intended SPC and NSE are equivalent. However, process-based greenhouse
purposes, or if it is suitable for use by others to serve their own goals. models are rarely based on linear regression. In fact, it is unclear why
Another possible reason for the creation of new models rather than SPC is used at all in this context: authors who use it state (sometimes
building on existing ones is an issue raised by Holzworth et al. (2015) in implicitly) that an SPC value close to 1 automatically represents good
the context of agricultural modelling: “maintenance of documentation model predictions. However, this is a misconception, as a high SPC will
and software/code has not been considered a core research outcome [...] only indicate that there is a linear relationship between measured and
This results in software that is maintained in an ad-hoc fashion to the predicted values. For instance, measured values of y = 1,2,3 and pre
point where often the best way forward in improving the software base dicted values of f = 500,0,-500 will yield SPC = 1, despite the model
is to start from scratch”. In this case, Holzworth et al. (2015) seem to being completely off in both the trend and the order of magnitude. These
assume that model code is available but poorly maintained. In green attributes led Bellocchi et al. (2010) to conclude that the use of the SPC
house modelling, matters are arguably worse, since it is rarely the case to evaluate model performance is “flawed”. Kobayashi and Salam (2000)
that code is made available at all, so that indeed the only way forward is give an example of how the SPC can be misleading, as it obscures
to start from scratch. important information for assessing model performance.
Reuse of existing models is also limited due to poor reporting. When Various scientific disciplines hold long ongoing debates regarding
code is not available, potential users are driven to reproduce models the most effective methods for model evaluation and validation (Bennett
based on equations printed in published papers. This method is et al., 2013; Eker et al., 2018; Legates and McCabe, 1999; Oreskes et al.,
vulnerable to mistakes such as misprints and omission of details, made 1994), including the crop modelling community (Bellocchi et al., 2010;
worse when old models are reprinted using new notation (e.g., new Cao et al., 2012; Kobayashi and Salam, 2000; Wallach et al., 2019a;
variable and parameter names). Another issue is imprecise referencing, Yang et al., 2014). It seems that in greenhouse modelling this debate has
for example when whole books (containing multiple models) are pro so far been absent, resulting in model developers each coming up with
vided as a reference to a model being used, or when several references their own methods for evaluation, and some simply not validating or
are given without an explanation on how they were used or combined. evaluating their model at all. The diversity in model validation metrics
The suspicion that new models are being developed simply because makes it difficult to compare validation results, which hinders
older models are not available is strengthened by the model inheritance straightforward model selection based on prediction accuracy. It is hard
chart (Fig. 5): when reusing or extending existing models, researchers to explain why so many metrics are being used, including some that
predominantly choose their own models as a basis for further work. were devised exclusively by the authors for a specific study. One possible
Cases where the work of others has been reused or extended are typically explanation is that authors choose metrics which they believe make
restricted to simple models, where reproduction is arguably easier. This their model look successful. Some evidence supporting this hypothesis is
observation is in line with that of Holzworth et al. (2015) who remarked: the fact that some validation results are presented in vague terms, e.g.
“[the] model fit-for-purpose question is usually overlooked in favor of “less than” or “greater than” some value (Table 4). Moreover, no
adopting an off-the-shelf model, likely one with which the researchers greenhouse model validation study ever seems to conclude that the
have some experience, regardless of its possible complexity misfit”. A model under investigation is poor, although positive results bias (Cata
similar tendency has also been found in the field of hydrological logue of Bias Collaboration, 2017) may also play a role here.
modelling (Addor and Melsen, 2019). The use of research facilities for evaluating energy use predictions
carries special difficulties, as research typically takes place in small
20
D. Katzin et al. Agricultural Systems 198 (2022) 103388
compartments within a bigger greenhouse. Edge effects due to a rela RMSE. Reporting on autocorrelation, bias, and other metrics of the error
tively large wall surface area, heat transport between compartments, is more nuanced than a simple yes or no answer to the question of
and a central heating system distributing heat to all compartments whether a model is valid, but it provides considerable insight to po
simultaneously, contribute error to the energy use measurements of tential users of the model. Naturally, publicly sharing the validation
individual compartments. These are all issues that are typically ignored data, including measurements, predictions, and record of control ac
in “perfectly stirred tank” type models, which often assume that the tions, would allow potential users to test the validation results using
greenhouse is infinitely large (Section 2.1). Ideally, authors should whatever metrics they like. However, sharing such data is extremely
report on the procedures taken to measure or estimate energy use in rare.
compartments, including what steps were taken to avoid edge effects. Similarly, it would be useful to evaluate model errors with respect to
Yield prediction was not part of the main focus of this study (Section the spatial variance that naturally exists in the greenhouse climate. As
2.1), but several greenhouse climate models reported on it during model shown by Van Beveren et al. (2015a), indoor air temperatures can differ
validation (Table 4). When evaluating yield predictions, one crucial by nearly 2 ◦ C, within the same horizontal plane.
component is the dry matter content (DMC) of the harvested product. At the same time, it should be stressed that models are a means to an
This parameter converts dry weight predictions – as they typically end, and the goal of achieving perfect model predictions is not very
appear in crop models – to fresh weight yield. In tomatoes, DMC ranges useful in and of itself. Model accuracy should be evaluated with respect
between 4 and 7.5% (Heuvelink et al., 2018), which results in a sub to the particular goal the model is meant to serve. It would be worth
stantial range of yield predictions given a particular dry weight while to reflect whether the model validation process properly serves
outcome: for 1 kg of dry weight, the range of 4–7.5% corresponds to a this evaluation. Moreover, assuming modelling studies are meant to
fresh weight prediction range of 13–25 kg. Thus, modifications within a eventually guide and shape horticultural practice, some reflection on the
generally accepted range of DMC can nearly double a given model's yield consequences of model uncertainty on actual practice would be useful.
prediction. Ideally, authors would report on how they estimate DMC For example, what is the practical meaning of inaccurate predictions of
when comparing dry weight model outputs to fresh weight yield indoor climate, yield, or energy use? A discussion on these consequences
measurements. would typically require some estimates on market prices, as well as
At the same time, there is room to consider how accurate we can other practical considerations such as resource availability. Neverthe
expect yield predictions for continuously yielding crops to really be. In less, such discussions would help carry modelling studies away from a
practice, growers have some flexibility regarding the moment of harvest purely theoretic realm, and place them within the context of the sector
(Saltveit, 2018), and harvesting strategy, influenced by concerns such as that they ideally should serve.
labor availability or produce price, may have a major influence on For instance, the acceptable error in models used for climate control
harvest (Marcelis and Gijzen, 1998). As long as such concerns are not might be very different from that of models used for system analysis.
included in greenhouse models, it may be wiser to settle for longer term Accurate, high frequency predictions of the indoor climate may not be
yield predictions (aggregated on a monthly or yearly basis) than to needed for models used in greenhouse structure design; for this goal,
expect accurate daily or weekly yield predictions. aggregation on bigger timescales is probably more suitable.
5.4.2. How accurate should process-based greenhouse climate models be? 5.4.3. Model parametrization
The lack of standards in greenhouse model validation raises the An important component of model development is setting the values
question how accurate greenhouse climate models really need to be. One for the model parameters. Parameter values may be based on previous
standard that began to establish regarding indoor greenhouse climate is literature, direct measurements, or calibration. Simple models typically
that a rRMSE of up to 10% is acceptable (e.g., Ahamed et al., 2018a; have relatively fewer parameters, but those tend to lump together
Sethi et al., 2013; Vanthoor et al., 2011b), although this choice lacks several processes in a way that is difficult to define or measure them
justification. In particular, while rRMSE provides a convenient com directly.
parison between predictions in different units, it can also be misleading Taking as an example two possible approaches for modelling tran
when comparing the same units under different settings. For example, spiration (Section 2.4.6), Eq. 12 uses only 2 parameters, but those must
an rRMSE of 10% corresponds to an RMSE of 1.5 ◦ C when the mean be calibrated for any specific setting by dedicated measurements that
temperature is 15 ◦ C but to an RMSE of 3 ◦ C when the mean temperature correlate radiation with crop transpiration. Conversely, Eq. 13 includes
is 30 ◦ C. The rRMSE is in fact unsuitable for non-absolute units such as 6 parameters, some of which are relatively easy to measure or to assume
◦
C: note that when the average temperature is 0 ◦ C, the rRMSE is un that they are within relatively narrow bands (e.g., the psychrometric
defined (or infinite). constant), while others are much more difficult to estimate or measure
In any case, it is counterproductive to set a fixed metric where directly (e.g., stomatal or aerodynamic conductance of a full crop).
models achieving values above or below some golden standard are Complex models tend to have more parameters, which increases the
deemed useful or useless. This lesson was painfully learned for the case risk of correlation between parameters, making calibration difficult. At
of statistical significance and p-values (Wasserstein and Lazar, 2016). the same time, parameters in complex models often express directly
Nevertheless, some guidelines may be useful here: a model cannot be measurable or known values. Simple models often include fewer pa
expected to be more accurate than the variance that is already present in rameters, but those must be obtained through calibration. Considering
the system, including measurement error. For instance, we cannot the fact that data used for calibration should typically be separate from
expect indoor climate predictions to be more accurate than the sensors the data used for validation (Wallach et al., 2019a), and the general
used for measurements. Greenhouse climate sensors often demonstrate scarcity of available data (Section 4.4), calibration of models, both
considerable measurement errors, which exceed the desired standards simple and complex, is often a challenging task.
set for the industry, even after sensor maintenance and calibration Nevertheless, there is room to consider the influence of any partic
(Bontsema et al., 2011). For example, the measurement error for indoor ular parameter on total model performance. A sensitivity analysis can
air temperature was found in one case to be in the range of 0.04–0.45 ◦ C help determine how a particular parameter influences model output,
(Bontsema et al., 2011). and which parameters should be calibrated or measured (Wallach et al.,
It would also be helpful if validation studies would provide infor 2019b). In any case, the objective of any given modelling study should
mation that helps distinguish between systematic and random errors in stand at the center of how parameters are chosen. A certain parameter
their predictions. For example, a model that provides accurate indoor may be extremely influential for a particular objective but almost
climate predictions, but with a time delay compared to the measured inconsequential for another. It is also important to clarify which
data, may be an extremely useful model but it would still produce a poor parametrization choices were made, and the motivations behind them.
21
D. Katzin et al. Agricultural Systems 198 (2022) 103388
Unfortunately, model developers often give very little information about redundant algorithms”. Unfortunately, this statement is still true for
this process. For examples, if direct measurements are performed to greenhouse climate models. Source code for greenhouse models is rarely
estimate parameters, it would be helpful to report not only on the mean made available by its authors, with a few exceptions emerging in the last
values obtained, but also on the parameter likelihood distribution. This years (Altes-Buch et al., 2019; Katzin et al., 2020b; Körner and Holst,
uncertainty in the parameter values could then be taken into account 2017). The models for Ahamed et al. (2018a, 2018b, 2018c, 2019) are
when evaluating the complete model. available in PDF form in Ahamed, 2018, providing transparency to the
studies but limiting reuse due to copyright. This state of affairs is a far
5.5. What advances in greenhouse modelling can we expect in the future? cry from that of agricultural models, where several long-standing
models have made their code available, albeit with differing levels of
5.5.1. Model development and extension accessibility (Soltani and Sinclair, 2015).
Ideally, and as already stated decades ago (Lacroix and Zanghi, 1990; Accessible model code would not only facilitate model reuse and
Van Bavel et al., 1985), more efforts would be placed on improving the extension, it could also help to compare different models, evaluate their
knowledge of individual processes in the greenhouse rather than validity in different scenarios, learn about the strengths and weaknesses
reproduce existing models. For example, considerable efforts are of each approach, and combine successful parts for further improve
currently being made in the modelling of CSGs. It would be worthwhile ment. In arable farming, such efforts have been ongoing in projects such
to test whether assumptions that were used when designing models for as the Agricultural Model Intercomparison and Improvement Project
Venlo-type greenhouses (e.g., an “infinite greenhouse”) still apply for (AgMIP) (Rosenzweig et al., 2013) and the Agricultural Model Exchange
CSG models. Similarly, there is a growing interest in modelling supple Initiative (AMEI) (Midingoyi et al., 2020).
mental lighting in greenhouses, as new technologies such as LEDs are An important component of model development and intercompar
rapidly advancing. Unfortunately, very few studies provide systemic ison is the development of benchmark datasets. Such datasets have been
measurements of lamp output and its influence on the greenhouse en made available for crop model validation and comparison (Asseng et al.,
ergy and other balances (Nelson and Bugbee (2015) is one rare 2015). Indeed, publicly available shared datasets would help evaluate
example). This issue is becoming increasingly important as plant phys models by comparing their prediction accuracy in equivalent scenarios.
iologists learn about the responses of crops to narrow-band LED spec Using such data, benchmark problems could be devised, setting quan
trums with the hopes of implementing these insights in greenhouses titative and qualitative standards that are expected of greenhouse
(Lazzarin et al., 2021; Ouzounis et al., 2015). climate models. Unfortunately, publicly available data from green
Other components of the greenhouse system which could use more houses is extremely scarce. Some exceptions are the Autonomous
development are mechanical cooling and dehumidification, where some Greenhouse Challenge (Hemming et al., 2019a, 2020) and the recently
advances have been made (De Zwart and Kempkes, 2008; Van Beveren established Controlled Environment Agriculture Open Data initiative
et al., 2015a; Vanthoor et al., 2011b) but complete model descriptions (CEAOD, 2020). Nevertheless, publicly available long-term data from
and validation studies are still rare. The crop is an important component large-scale commercial greenhouses is currently missing.
of the greenhouse climate, both for estimating photosynthesis and the
system performance in terms of yield. In this respect, there remains a 5.5.3. Model selection
need for reliable quantitative information. For example, quantitative The models explored in this study are in their core essentially the
knowledge on the influence of adverse climate on crop behavior is same. All models describe similar key variables, and include processes
limited (Vanthoor et al., 2011a). There is also very little quantitative derived from basic laws of physics concerning the conservation of mass
knowledge on the influence of the indoor climate on the occurrence and and energy. Nevertheless, the models differ in the details of the indi
spread of disease, although some work has been done, particularly vidual parameters and transport phenomena described. Ideally, the
regarding the prevention of botrytis (Baptista, 2007; Baptista et al., choice of these details should depend on a specific study's objectives and
2012; Cañadas et al., 2017; Körner et al., 2014). Furthermore, models goals. Unfortunately, the reasoning behind the choices made during
describing the influence of irrigation water on the crop (e.g., Jiang et al., model development, or a thorough analysis of the pros and cons of these
2019) are rarely included in greenhouse models. In particular, dry choices, is largely left undocumented.
matter content (DMC) remains an elusive but extremely important The overview given in this study could help greenhouse modelers
variable with very few studies attempting to accurately predict it. Other reflect on the issues faced in model selection and development, and
underdeveloped components in greenhouse crop models were noted by consider the range of suitability, complexity, validity, and transparency
Marcelis et al. (1998) and include product quality, leaf area develop of their model. While the models included in this review are only a small
ment, maintenance respiration and organ abortion. part of the vast range of greenhouse climate models that have been
An interesting feature of practically all detailed process-based developed, this study provides an initial point of reference and directs
greenhouse models is the fact that stomatal conductance is modelled researchers to consider several aspects, both when choosing an existing
separately for transpiration and photosynthesis. Stomata simultaneously model to build on and when presenting their own work.
regulate the exchange of water vapor and CO2 between the air and the An alternative approach to model selection is the use of multimodel
crop, and this principle lies at the basis of stomatal modelling (Buckley, ensembles. In this approach, multiple models are used in parallel to
2017). This principle is often overlooked in greenhouse modelling. On predict a certain outcome. The range, variability, and average of the
the one hand this may be expected, as many models do not include a CO2 model predictions can provide valuable insights on the system being
balance (Table 3). But even when the CO2 balance is absent, many modelled and on the models used to predict its performance. This
process-based transpiration models are based on a prediction of stomatal approach has been tried in the context of crop modelling, producing
behavior (Katsoulas and Stanghellini, 2019). For these models, it would some promising results (Martre et al., 2015; Wallach et al., 2018).
be informative to test how the modelled stomatal behavior influences When developing new models, developers should explore which
photosynthesis and crop growth, especially in models that have detailed models already exist that describe the system they are interested in, such
descriptions of both transpiration and photosynthesis (e.g., Golzar et al., as the type of greenhouse, crop, and equipment included in their system.
2018; Vanthoor et al., 2011a; Vanthoor et al., 2011b). Developers should consider the objectives of previous models and reflect
critically whether using a given model is suitable to satisfy their own
5.5.2. Model transparency and comparison objectives. Model validation studies are useful for demonstrating
Forty years ago, in the context of geographical models, Willmott whether a given model accurately represents a particular system. At the
(1981) lamented that “far too few computer programs have been pub same time, the structure and assumptions underlying the model are
lished, resulting in the development of numerous overlapping and/or equally important for judging the suitability of a model for a particular
22
D. Katzin et al. Agricultural Systems 198 (2022) 103388
23
D. Katzin et al. Agricultural Systems 198 (2022) 103388
being and energy needs. Appl. Therm. Eng. 179, 115698 https://doi.org/10.1016/j. De Zwart, H.F., Kempkes, F., 2008. Characterizing of cooling equipment for closed
applthermaleng.2020.115698. greenhouses. Acta Hortic. 801, 409–415. https://doi.org/10.17660/
Baptista, F.J., 2007. Modelling the Climate in Unheated Tomato Greenhouses and actahortic.2008.801.43.
Predicting Botrytis Cinerea Infection. Universidade de Évora. Draoui, B., 1994. Caractérisation et analyse du comportement thermohydrique d’une
Baptista, F.J., Bailey, B.J., Meneses, J.F., 2012. Effect of nocturnal ventilation on the serre horticole. L’université de Nice Sophia Antipolis.
occurrence of Botrytis cinerea in Mediterranean unheated tomato greenhouses. Crop Duncan, W.G., Maize, L., Evans, T., 1975. Crop Physiology - some Case Histories 374.
Prot. 32, 144–149. https://doi.org/10.1016/j.cropro.2011.11.005. Cambridge University Press.
Bellocchi, G., Rivington, M., Donatelli, M., Matthews, K., 2010. Validation of biophysical Eker, S., Rovenskaya, E., Obersteiner, M., Langan, S., 2018. Practice and perspectives in
models: issues and methodologies. A review. Agronomy for Sustainable Development the validation of resource management models. Nat. Commun. 9 (1), 1–10. https://
30 (1), 109–130. https://doi.org/10.1051/agro/2009001. doi.org/10.1038/s41467-018-07811-9.
Bennett, N.D., Croke, B.F.W., Guariso, G., Guillaume, J.H.A., Hamilton, S.H., Esmaeli, H., Roshandel, R., 2017. Thermal model development for solar greenhouse
Jakeman, A.J., Marsili-Libelli, S., Newham, L.T.H., Norton, J.P., Perrin, C., Pierce, S. considering climate condition. Proceedings of 1–12. https://doi.org/10.18086/
A., Robson, B., Seppelt, R., Voinov, A.A., Fath, B.D., Andreassian, V., 2013. swc.2017.26.04. SWC2017/SHC2017.
Characterising performance of environmental models. Environ. Model. Softw. 40, Esmaeli, H., Roshandel, R., 2020. Optimal design for solar greenhouses based on climate
1–20. https://doi.org/10.1016/j.envsoft.2012.09.011. conditions. Renew. Energy 145, 1255–1265. https://doi.org/10.1016/j.
Blasco, X., Martínez, M., Herrero, J.M., Ramos, C., Sanchis, J., 2007. Model-based renene.2019.06.090.
predictive control of greenhouse climate for reducing energy and water Evangelisti, L., Guattari, C., Asdrubali, F., 2019. On the sky temperature models and their
consumption. Comput. Electron. Agric. 55 (1), 49–70. https://doi.org/10.1016/j. influence on buildings energy performance: A critical review. Energy and Buildings
compag.2006.12.001. 183, 607–625. https://doi.org/10.1016/J.ENBUILD.2018.11.037.
B-Mex, 2020. About B-Mex. https://b-mex.nl/OverOns.html. Forrester, J.W., 1961. Industrial Dynamics. MIT Press.
Bontsema, J., Hemming, J., Stanghellini, C., De Visser, P.H.B., Van Henten, E.J., Gary, C., Jones, J.W., Tchamitchian, M., 1998. Crop modelling in horticulture: state of
Budding, J., Rieswijk, T., Nieboer, S., 2007. On-line estimation of the transpiration in the art. Sci. Hortic. 74 (1), 3–20. https://doi.org/10.1016/S0304-4238(98)00080-6.
greenhouse horticulture. Proceedings Agricontrol 3–5. Garzoli, K., 1985. A simple greenhouse climate model. In. Acta Hortic. 174, 393–400.
Bontsema, J., Van Henten, E.J., Gieling, T.H., Swinkels, G.L.A.M., 2011. The effect of Gharghory, S.M., 2020. Deep network based on long short-term memory for time series
sensor errors on production and energy consumption in greenhouse horticulture. prediction of microclimate data inside the greenhouse. Int. J. Comput. Intell. Appl.
Comput. Electron. Agric. 79 (1), 63–66. https://doi.org/10.1016/j. 19 (02), 2050013. https://doi.org/10.1142/S1469026820500133.
compag.2011.08.008. Golzar, F., Heeren, N., Hellweg, S., Roshandel, R., 2018. A novel integrated framework to
Bot, G.P.A., 1983. Greenhouse Climate: From Physical Processes to a Dynamic Model. evaluate greenhouse energy demand and crop yield production. Renew. Sust. Energ.
Boulard, T., Baille, A., 1987. Analysis of thermal performance of a greenhouse as a solar Rev. 96, 487–501. https://doi.org/10.1016/j.rser.2018.06.046.
collector. Energy in Agriculture 6 (1), 17–26. https://doi.org/10.1016/0167-5826 Golzar, F., Heeren, N., Hellweg, S., Roshandel, R., 2019. A comparative study on the
(87)90018-0. environmental impact of greenhouses: A probabilistic approach. Sci. Total Environ.
Boulard, T., Draoui, B., Neirac, F., 1996. Calibration and validation of a greenhouse 675, 560–569. https://doi.org/10.1016/j.scitotenv.2019.04.092.
climate control model. Acta Horticulturae 406, 49–61. https://doi.org/10.17660/ Goudriaan, J., Van Laar, H.H., 1993. Modelling potential crop growth processes:
actahortic.1996.406.4. Textbook with exercises. In: Current Issues in Production Ecology, vol. 2. Kluwer
Boulard, T., Kittas, C., Roy, J.C., Wang, S., 2002. Convective and ventilation transfers in Academic Publishers. https://doi.org/10.1017/CBO9781107415324.004.
greenhouses, part 2: determination of the distributed greenhouse climate. Biosyst. Haefner, J.W., 2005. Qualitative model formation. In: Modeling Biological Systems:
Eng. 83 (2), 129–147. https://doi.org/10.1006/bioe.2002.0114. Principles and Applications, 2nd ed. Springer Science+Business Media.
Buckley, T.N., 2017. Modeling stomatal conductance. Plant Physiol. 174 (June) https:// Hammer, G., Messina, C., Wu, A., Cooper, M., 2019. Biological reality and parsimony in
doi.org/10.1104/pp.16.01772. crop models—why we need both in crop improvement! In Silico Plants 1 (1). https://
Businger, J.A., 1963. The glasshouse (greenhouse) climate. In: Van Wijk, W.R. (Ed.), doi.org/10.1093/insilicoplants/diz010.
Physics of Plant Environment. North-Holland, pp. 277–318. Hasni, A., Taibi, R., Draoui, B., Boulard, T., 2011. Optimization of greenhouse climate
Cañadas, J., Sánchez-Molina, J.A., Rodríguez, F., del Águila, I.M., 2017. Improving model parameters using particle swarm optimization and genetic algorithms. Energy
automatic climate control with decision support techniques to minimize disease Procedia 6, 371–380. https://doi.org/10.1016/j.egypro.2011.05.043.
effects in greenhouse tomatoes. Information Processing in Agriculture 4 (1), 50–63. Hemming, S., De Zwart, H.F., Elings, A., Righini, I., Petropoulou, A., 2019a. Autonomous
https://doi.org/10.1016/J.INPA.2016.12.002. greenhouse challenge, first edition (2018). 4TU.Centre for Research Data. Dataset.
Cao, H.X., Hanan, J.S., Liu, Y., Liu, Y.X., Yue, Y. Bin, Zhu, D.W., Lu, J.F., Sun, J.Y., Shi, C. https://doi.org/10.4121/uuid:e4987a7b-04dd-4c89-9b18-883aad30ba9a.
L., Ge, D.K., Wei, X.F., Yao, A.Q., Tian, P.P., Bao, T.L., 2012. Comparison of crop Hemming, S., De Zwart, H.F., Elings, A., Righini, I., Petropoulou, A., 2019b. Remote
model validation methods. J. Integr. Agric. 11 (8), 1274–1285. https://doi.org/ control of greenhouse vegetable production with artificial intelligence—greenhouse
10.1016/S2095-3119(12)60124-5. climate, irrigation, and crop production. Sensors 19 (8), 1807. https://doi.org/
Catalogue of Bias Collaboration, Plüddemann, A., Banerjee, A., O’Sullivan, J., 2017. 10.3390/s19081807.
Positive results bias. Catalogue Of Biases. https://www.catalogueofbiases.org/biases Hemming, S., De Zwart, H.F., Elings, A., Petropoulou, A., Righini, I., 2020. Autonomous
/positive-results-bias. greenhouse challenge, second edition (2019). 4TU.ResearchData. https://doi.org/
CEAOD, 2020. Controlled Environment Agriculture Open Data. https://ceaod.github.io/. 10.4121/uuid:88d22c60-21b3-4ea8-90db-20249a5be2a7.
Chen, J., Zhao, J., Xu, F., Hu, H., Ai, Q., Yang, J., 2015. Modeling of energy demand in Heuvelink, E., Li, T., Dorais, M., 2018. Crop growth and yield. In: Heuvelink, E. (Ed.),
the greenhouse using PSO-GA hybrid algorithms. Math. Probl. Eng. 2015 https://doi. Tomatoes, 2nd ed. Cabi, pp. 89–135. https://doi.org/10.1079/
org/10.1155/2015/871075. 9781780641935.0000.
Chen, J., Yang, J., Zhao, J., Xu, F., Shen, Z., Zhang, L., 2016. Energy demand forecasting Hikosaka, K., Niinemets, Ü., Anten, N.P.R., 2016. Canopy photosynthesis: from basics to
of the greenhouses using nonlinear models based on model optimized prediction applications. In: Hikosaka, K., Niinemets, Ü., Anten, N.P.R. (Eds.), Canopy
method. Neurocomputing 174, 1087–1100. https://doi.org/10.1016/j. Photosynthesis: From Basics to Applications SE - 6, vol. 42. https://doi.org/10.1007/
neucom.2015.09.105. 978-94-017-7291-4_6.
Chen, J., Xu, F., Ding, B., Wu, N., Shen, Z., Zhang, L., 2019a. Performance analysis of Hölscher, T., 1992. Ein simulationsmodell für das gewächshausklima. University of
radiation and electricity yield in a photovoltaic panel integrated greenhouse using Hannover.
the radiation and thermal models. Comput. Electron. Agric. 164 https://doi.org/ Holzworth, D.P., Snow, V., Janssen, S., Athanasiadis, I.N., Donatelli, M.,
10.1016/j.compag.2019.104904. Hoogenboom, G., White, J.W., Thorburn, P., 2015. Agricultural production systems
Chen, C., Yu, N., Yang, F., Mahkamov, K., Han, F., Li, Y., Ling, H., 2019b. Theoretical and modelling and software: current status and future prospects. Environ. Model Softw.
experimental study on selection of physical dimensions of passive solar greenhouses 72, 276–286. https://doi.org/10.1016/j.envsoft.2014.12.013.
for enhanced energy performance. Sol. Energy 191, 46–56. https://doi.org/ Hoogendoorn Growth Management, 2020. Harvest Forecast. https://www.hoogendoorn.
10.1016/j.solener.2019.07.089. nl/en/product/harvest-forecast/.
Choab, N., Allouhi, A., El Maakoul, A., Kousksou, T., Saadeddine, S., Jamil, A., 2019. Iddio, E., Wang, L., Thomas, Y., McMorrow, G., Denzer, A., 2020. Energy efficient
Review on greenhouse microclimate and application: design parameters, thermal operation and modeling for greenhouses: A literature review. Renew. Sust. Energ.
modeling and simulation, climate controlling technologies. Sol. Energy 191, Rev. 117, 109480 https://doi.org/10.1016/j.rser.2019.109480.
109–137. https://doi.org/10.1016/j.solener.2019.08.042. Janssen, S.J.C., Porter, C.H., Moore, A.D., Athanasiadis, I.N., Foster, I., Jones, J.W.,
Crout, N.M.J., Tarsitano, D., Wood, A.T., 2009. Is my model too complex? Evaluating Antle, J.M., 2017. Towards a new generation of agricultural system data, models and
model formulation using model reduction. Environ. Model. Softw. 24 (1), 1–7. knowledge products: information and communication technology. Agric. Syst. 155,
https://doi.org/10.1016/j.envsoft.2008.06.004. 200–212. https://doi.org/10.1016/j.agsy.2016.09.017.
De Halleux, D., 1989. Modèle dynamique des échanges énergétiques des serres: étude Jarvis, P.G., 1976. The interpretation of the variations in leaf water potential and
théorique et expérimentale. Gembloux, Belgium. stomatal conductance found in canopies in the field. Philosophical Transactions of
De Ridder, F., Van Roy, J., Vanlommel, W., Van Calenberge, B., Vliex, M., De Win, J., De the Royal Society of London. B 273, 593–610. https://doi.org/10.1098/
Schutter, B., Binnemans, S., De Pauw, M., 2020. Convex parameter estimator for RSTB.1976.0035.
grey-box models, applied to characterise heat flows in greenhouses. Biosyst. Eng. Jiang, X., Zhao, Y., Wang, R., Zhao, S., 2019. Modeling the relationship of tomato yield
191, 13–26. https://doi.org/10.1016/j.biosystemseng.2019.12.009. parameters with deficit irrigation at different growth stages. HortScience 54 (9),
De Zwart, H.F., 1993. Determination of direct transmission of a multispan greenhouse 1492–1500. https://doi.org/10.21273/HORTSCI14179-19.
using vector algebra. J. Agric. Eng. Res. 56, 39–49. https://doi.org/10.1006/ Jolliet, O., Danloy, L., Gay, J.B., Munday, G.L., Reist, A., 1991. HORTICERN: an
jaer.1993.1059. improved static model for predicting the energy consumption of a greenhouse. Agric.
De Zwart, H.F., 1996. Analyzing Energy-Saving Options in Greenhouse Cultivation Using For. Meteorol. 55 (3–4), 265–294. https://doi.org/10.1016/0168-1923(91)90066-Y.
a Simulation Model. Landbouwuniversiteit, Wageningen.
24
D. Katzin et al. Agricultural Systems 198 (2022) 103388
Jomaa, M., Tadeo, F., Abbes, M., Abdelkader, M., 2019. Greenhouse modeling, validation many programming languages and simulation platforms. In Silico Plants. https://
and climate control based on fuzzy logic. Engineering, Technology & Applied doi.org/10.1093/insilicoplants/diaa007.
Science Research 9 (4), 4405–4410. Mohamed, S., Hameed, I.A., 2018. A GA-based adaptive neuro-fuzzy controller for
Jones, J.W., Kenig, A., Vallejos, C.E., 1999. Reduced state-variable tomato growth model. greenhouse climate control system. Alexandria Engineering Journal 57, 773–779.
Trans. Am. Soc. Agric. Eng. 42 (1994), 255–265. https://doi.org/10.1016/j.aej.2014.04.009.
Katsoulas, N., Stanghellini, C., 2019. Modelling crop transpiration in greenhouses: Mohammadi, B., Ranjbar, F., Ajabshirchi, Y., 2020. Exergoeconomic analysis and multi-
different models for different applications. Agronomy 9 (7), 392. https://doi.org/ objective optimization of a semi-solar greenhouse with experimental validation.
10.3390/agronomy9070392. Appl. Therm. Eng. 164, 114563 https://doi.org/10.1016/j.
Katzin, D., Van Mourik, S., Kempkes, F., Van Henten, E.J., 2020a. Energy saving applthermaleng.2019.114563.
measures in optimally controlled greenhouse lettuce cultivation. Acta Hortic. 1271, Monteith, J.L., 1965. Evaporation and environment. Symposia of the Society for
265–272. https://doi.org/10.17660/ActaHortic.2020.1271.36. Experimental Biology. Symposia of the Society for Experimental Biology 19,
Katzin, D., Van Mourik, S., Kempkes, F., Van Henten, E.J., 2020b. GreenLight – an open 205–234.
source model for greenhouses with supplemental lighting: evaluation of heat Monteith, J.L., 1996. The quest for balance in crop modeling. Agron. J. 88, 695–697.
requirements under LED and HPS lamps. Biosyst. Eng. 194, 61–81. https://doi.org/ Morris, L.G., Trickett, E.S., Vanstone, F.H., Wells, D.A., 1958. The limitation of maximum
10.1016/j.biosystemseng.2020.03.010. temperature in a glasshouse by the use of a water film on the roof. J. Agric. Eng. Res.
Keating, B.A., 2020. Crop, soil and farm systems models – science, engineering or snake 3, 121–130.
oil revisited. Agric. Syst. 184, 102903 https://doi.org/10.1016/j.agsy.2020.102903. Nelson, J.A., Bugbee, B., 2014. Economic analysis of greenhouse lighting: light emitting
Keating, B.A., Thorburn, P.J., 2018. Modelling crops and cropping systems - evolving diodes vs. High Intensity Discharge Fixtures. PLoS ONE 9 (6), e99010. https://doi.
purpose, practice and prospects. Eur. J. Agron. 100, 163–176. https://doi.org/ org/10.1371/journal.pone.0099010.
10.1016/j.eja.2018.04.007. Nelson, J.A., Bugbee, B., 2015. Analysis of environmental effects on leaf temperature
Kobayashi, K., Salam, M.U., 2000. Comparing simulated and measured values using under sunlight, high pressure sodium and light emitting diodes. PLoS One 10 (10),
mean squared deviation and its components. Agron. J. 92 (March), 345–352. e0138930. https://doi.org/10.1371/journal.pone.0138930.
https://doi.org/10.1007/s100870050043. Norton, T., Sun, D.-W., Grant, J., Fallon, R., Dodd, V., 2007. Applications of
Körner, O., 2019. Models, sensors and decision support systems in greenhouse computational fluid dynamics (CFD) in the modelling and design of ventilation
cultivation. In: Marcelis, L.F.M., Heuvelink, E. (Eds.), Achieving Sustainable systems in the agricultural industry: A review. Bioresour. Technol. 98 (12),
Greenhouse Production, 1st ed. Burleigh Dodds Science Publishing, pp. 379–412. 2386–2414. https://doi.org/10.1016/j.biortech.2006.11.025.
https://doi.org/10.19103/AS.2019.0052.15. Oreskes, N., Shrader-Frechette, K., Belitz, K., 1994. Verification, validation, and
Körner, O., Holst, N., 2017. An open-source greenhouse modelling platform. Acta Hortic. confirmation of numerical models in the earth sciences. Science 263 (5147),
1154, 241–248. https://doi.org/10.17660/ActaHortic.2017.1154.32. 641–646.
Körner, O., Holst, N., De Visser, P.H.B., 2014. A model-based decision support tool for Ouzounis, T., Rosenqvist, E., Ottosen, C.O., 2015. Spectral effects of artificial light on
grey mould prediction. Acta Hortic. 1037, 569–574. https://doi.org/10.17660/ plant physiology and secondary metabolism: A review. HortScience 50 (8),
ActaHortic.2014.1037.71. 1128–1135.
Kuijpers, W.J.P., Van de Molengraft, M.J.G., Van Mourik, S., Van’t Ooster, A., Pasgianos, G.D., Arvanitis, K.G., Polycarpou, P., Sigrimis, N., 2003. A nonlinear feedback
Hemming, S., Van Henten, E.J., 2019. Model selection with a common structure: technique for greenhouse environmental control. Comput. Electron. Agric. 40,
tomato crop growth models. Biosyst. Eng. 187, 247–257. https://doi.org/10.1016/j. 153–177. https://doi.org/10.1016/S0168-1699(03)00018-8.
biosystemseng.2019.09.010. Passioura, J., 1973. Sense and nonsense in crop simulation. J. Aust. Inst. Agric. Sci. 39
Kuijpers, W.J.P., Katzin, D., Van Mourik, S., Antunes, D.J., Hemming, S., Van de (3), 181–183.
Molengraft, M.J.G., 2021. Lighting systems and strategies compared in an optimally Passioura, J., 1996. Simulation models: science, snake oil, education, or engineering?
controlled greenhouse. Biosyst. Eng. 202, 195–216. https://doi.org/10.1016/j. Agron. J. 88, 690–694.
biosystemseng.2020.12.006. Pérez-González, A., Begovich-Mendoza, O., Ruiz-León, J., 2018. Modeling of a
Lacroix, R., Zanghi, J.C., 1990. Étude comparative de la structure des modèles de greenhouse prototype using PSO and differential evolution algorithms based on a
transfert d’énergie et de masse dans les serres. Can. Agric. Eng. 32, 269–284. real-time LabViewTM application. Applied Soft Computing Journal 62, 86–100.
Lammari, K., Bounaama, F., Draoui, B., Mrah, B., Haidas, M., 2012. GA optimization of https://doi.org/10.1016/j.asoc.2017.10.023.
the coupled climate model of an order two of a greenhouse. Energy Procedia 18, Piscia, D., Muñoz, P., Panadès, C., Montero, J.I., 2015. A method of coupling CFD and
416–425. https://doi.org/10.1016/j.egypro.2012.05.053. energy balance simulations to study humidity control in unheated greenhouses.
Lammari, K., Bounaama, F., Ouradj, B., Draoui, B., 2020. Constrained GA PI sliding mode Comput. Electron. Agric. 115, 129–141. https://doi.org/10.1016/j.
control of indoor climate coupled MIMO greenhouse model. Journal of Thermal compag.2015.05.005.
Engineering 6 (3), 313–326. https://doi.org/10.18186/THERMAL.711554. Priva, 2019. Plantonomy helps big greenhouses scale up and give small greenhouses
Lazzarin, M., Meisenburg, M., Meijer, D., van Ieperen, W., Marcelis, L.F.M., Kappers, I.F., stability. https://www.meetphil.com/insights/plantonomy-helps-big-greenhouses-s
van der Krol, A.R., van Loon, J.J.A., Dicke, M., 2021. LEDs make it resilient: effects cale-up-and-give-small-greenhouses-stability.
on plant growth and defense. Trends Plant Sci. 26 (5), 496–508. https://doi.org/ Rabobank, 2018. World Vegetable Map 2018. https://research.rabobank.com/far/en/se
10.1016/j.tplants.2020.11.013. ctors/regional-food-agri/world_vegetable_map_2018.html.
Legates, D.R., McCabe, G.J., 1999. Evaluating the use of “goodness-of-fit” measures in Rasheed, A., Na, W.H., Lee, J.W., Kim, H.T., Lee, H.W., 2019. Optimization of
hydrologic and hydroclimatic model validation. Water Resour. Res. 35 (1), 233–241. greenhouse thermal screens for maximized energy conservation. Energies 12 (3592).
https://doi.org/10.1029/1998WR900018. https://doi.org/10.3390/en12193592.
Lentz, W., 1998. Model applications in horticulture: a review. Sci. Hortic. 74 (1), Righini, I., Vanthoor, B., Verheul, M.J., Naseer, M., Maessen, H., Persson, T.,
151–174. https://doi.org/10.1016/S0304-4238(98)00085-5. Stanghellini, C., 2020. A greenhouse climate-yield model focussing on additional
Lopez-Cruz, I.L., Fitz-Rodríguez, E., Salazar-Moreno, R., Rojano-Aguilar, A., Kacira, M., light, heat harvesting and its validation. Biosyst. Eng. 194, 1–15. https://doi.org/
2018. Development and analysis of dynamical mathematical models of greenhouse 10.1016/j.biosystemseng.2020.03.009.
climate : A review. Eur. J. Hortic. Sci. 83 (5), 269–280. Rosenzweig, C., Jones, J.W., Hatfield, J.L., Ruane, A.C., Boote, K.J., Thorburn, P.,
Ma, D., Carpenter, N., Maki, H., Rehman, T.U., Tuinstra, M.R., Jin, J., 2019. Greenhouse Antle, J.M., Nelson, G.C., Porter, C., Janssen, S., Asseng, S., Basso, B., Ewert, F.,
environment modeling and simulation for microclimate control. Comput. Electron. Wallach, D., Baigorria, G., Winter, J.M., 2013. The agricultural model
Agric. 162, 134–142. https://doi.org/10.1016/j.compag.2019.04.013. Intercomparison and improvement project (AgMIP): protocols and pilot studies.
Marcelis, L.F.M., Gijzen, H., 1998. Evaluation under commercial conditions of a model of Agric. For. Meteorol. 170, 166–182. https://doi.org/10.1016/j.
prediction of the yield and quality of cucumber fruits. Sci. Hortic. 76, 171–181. agrformet.2012.09.011.
https://doi.org/10.1016/S0304-4238(98)00156-3. Roy, J.C., Boulard, T., Kittas, C., Wang, S., 2002. Convective and ventilation transfers in
Marcelis, L.F.M., Heuvelink, E., Goudriaan, J., 1998. Modelling biomass production and greenhouses, part 1: the greenhouse considered as a perfectly stirred tank. Biosyst.
yield of horticultural crops: A review. Sci. Hortic. 74 (1–2), 83–111. https://doi.org/ Eng. 83 (1), 1–20. https://doi.org/10.1006/bioe.2002.0107.
10.1016/S0304-4238(98)00083-1. Saltveit, M.E., 2018. Postharvest biology and handling of tomatoes. In: Heuvelink, E.
Marcelis, L.F.M., Elings, A., De Visser, P.H.B., Heuvelink, E., 2009. Simulating growth (Ed.), Tomatoes, 2nd ed. Cabi, pp. 314–336. https://doi.org/10.1079/
and development of tomato crop. Acta Hortic. 821, 101–110. https://doi.org/ 9781780641935.0000.
10.17660/ActaHortic.2009.821.10. Seginer, I., Van Beveren, P.J.M., Van Straten, G., 2018. Day-to-night heat storage in
Marcelis, L.F.M., Costa, J.M., Heuvelink, E., 2019. Achieving sustainable greenhouse greenhouses: 3 co-generation of heat and electricity (CHP). Biosyst. Eng. 172, 1–18.
production: present status, recent advances and future developments. In: Marcelis, L. https://doi.org/10.1016/j.biosystemseng.2018.05.006.
F.M., Heuvelink, E. (Eds.), Achieving Sustainable Greenhouse Production. Burleigh Seginer, I., Van Straten, G., Van Beveren, P.J.M., 2020a. Modelling evapotranspiration in
Dodds Science Publishing, pp. 1–14. https://doi.org/10.19103/AS.2019.0052.01. off-line simulations of greenhouse climate control. Acta Hortic. 1296, 341–348.
Martre, P., Wallach, D., Asseng, S., Ewert, F., Jones, J.W., Rötter, R.P., Boote, K.J., https://doi.org/10.17660/actahortic.2020.1296.44.
Ruane, A.C., Thorburn, P.J., Cammarano, D., Hatfield, J.L., Rosenzweig, C., Seginer, I., Van Straten, G., Van Beveren, P.J.M., 2020b. Day-to-night heat storage in
Aggarwal, P.K., Angulo, C., Basso, B., Bertuzzi, P., Biernath, C., Brisson, N., greenhouses: 4. Changing the environmental bounds. Biosystems Engineering 192,
Challinor, A.J., Wolf, J., 2015. Multimodel ensembles of wheat growth: many models 90–107. https://doi.org/10.1016/j.biosystemseng.2020.01.005.
are better than one. Glob. Chang. Biol. 21 (2), 911–925. https://doi.org/10.1111/ Sethi, V.P., 2019. Thermal modelling of asymmetric overlap roof greenhouse with
gcb.12768. experimental validation. International Journal of Sustainable Energy 38 (1), 24–36.
Midingoyi, C.A., Pradal, C., Athanasiadis, I.N., Donatelli, M., Enders, A., Fumagalli, D., https://doi.org/10.1080/14786451.2018.1424167.
Garcia, F., Holzworth, D., Hoogenboom, G., Porter, C., Raynal, H., Thorburn, P., Sethi, V.P., Sumathy, K., Lee, C., Pal, D.S., 2013. Thermal modeling aspects of solar
Martre, P., 2020. Reuse of process-based models: automatic transformation into greenhouse microclimate control: A review on heating technologies. Sol. Energy 96,
56–82. https://doi.org/10.1016/j.solener.2013.06.034.
25
D. Katzin et al. Agricultural Systems 198 (2022) 103388
Sinclair, T.R., Seligman, N., 2000. Criteria for publishing papers on crop modeling. Field Van Henten, E.J., Bontsema, J., 2009. Time-scale decomposition of an optimal control
Crop Res. 68 (3), 165–172. https://doi.org/10.1016/S0378-4290(00)00105-2. problem in greenhouse climate management. Control. Eng. Pract. 17 (1), 88–96.
Soltani, A., Sinclair, T.R., 2015. A comparison of four wheat models with respect to https://doi.org/10.1016/j.conengprac.2008.05.008.
robustness and transparency: simulation in a temperate, sub-humid environment. Van Henten, E.J., Bontsema, J., Van Straten, G., 1997. Improving the efficiency of
Field Crop Res. 175, 37–46. https://doi.org/10.1016/j.fcr.2014.10.019. greenhouse climate control: an optimal control approach. Neth. J. Agric. Sci. 45,
Stanghellini, C., 1987. Transpiration of Greenhouse Crops: An Aid to Climate 109–125.
Management. Landbouwuniversiteit Wageningen. Van Mourik, S., 2008. Modelling and Control of Systems with Flow. University of
Stanghellini, C., Bontsema, J., De Koning, A., Baeza, E.J., 2012. An algorithm for optimal Twente. https://doi.org/10.3990/1.9789036526173.
fertilization with pure carbon dioxide in greenhouses. Acta Hortic. 952, 119–124. Van Ooteghem, R.J.C., 2007. Optimal Control Design for a Solar Greenhouse.
https://doi.org/10.17660/ActaHortic.2012.952.13. Wageningen University.
Stanghellini, C., Van’t Ooster, A., Heuvelink, E., 2019. Greenhouse Horticulture: Van Straten, G., Van Willigenburg, G., Van Henten, E.J., Van Ooteghem, R.J.C., 2010.
Technology for Optimal Crop Production, 1st ed. Wageningen Academic Publishers. Optimal Control of Greenhouse Cultivation. CRC Press.
https://doi.org/10.3920/978-90-8686-879-7. Vanthoor, B., 2011. A Model Based Greenhouse Design Method. Wageningen University.
Stockle, C.O., 1992. Canopy photosynthesis and transpiration estimates using radiation Vanthoor, B., De Visser, P.H.B., Stanghellini, C., Van Henten, E.J., 2011a. A methodology
interception models with different levels of detail. Ecol. Model. 60 (1), 31–44. for model-based greenhouse design: part 2, description and validation of a tomato
https://doi.org/10.1016/0304-3800(92)90011-3. yield model. Biosyst. Eng. 110 (4), 378–395. https://doi.org/10.1016/j.
Su, Y., Xu, L., Goodman, E.D., 2018. Nearly dynamic programming NN-approximation- biosystemseng.2011.08.005.
based optimal control for greenhouse climate: A simulation study. Optimal Control Vanthoor, B., Stanghellini, C., Van Henten, E.J., De Visser, P.H.B., 2011b. A methodology
Applications and Methods 39 (2), 638–662. https://doi.org/10.1002/oca.2370. for model-based greenhouse design: part 1, a greenhouse climate model for a broad
Subin, M.C., Singh, A., Kalaichelvi, V., Karthikeyan, R., Periasamy, C., 2020. Design and range of designs and climates. Biosyst. Eng. 110 (4), 363–377. https://doi.org/
robustness analysis of intelligent controllers for commercial greenhouse. Mechanical 10.1016/j.biosystemseng.2011.06.001.
Sciences 11 (2), 299–316. https://doi.org/10.5194/ms-11-299-2020. Vanthoor, B., Van Henten, E.J., Stanghellini, C., De Visser, P.H.B., 2011c. A methodology
Taki, M., Ajabshirchi, Y., Ranjbar, S.F., Rohani, A., Matloobi, M., 2016. Modeling and for model-based greenhouse design: part 3, sensitivity analysis of a combined
experimental validation of heat transfer and energy consumption in an innovative greenhouse climate-crop yield model. Biosyst. Eng. 110 (4), 396–412. https://doi.
greenhouse structure. Information Processing in Agriculture 3 (3), 157–174. https:// org/10.1016/j.biosystemseng.2011.08.006.
doi.org/10.1016/j.inpa.2016.06.002. Villarreal-Guerrero, F., Kacira, M., Fitz-Rodríguez, E., Kubota, C., Giacomelli, G.A.,
Taki, M., Rohani, A., Rahmati-Joneidabad, M., 2018. Solar Thermal Simulation and Linker, R., Arbel, A., 2012. Comparison of three evapotranspiration models for a
Applications in Greenhouse. In: Information Processing in Agriculture, vol. 5. China greenhouse cooling strategy with natural ventilation and variable high pressure
Agricultural University, pp. 83–113. https://doi.org/10.1016/j.inpa.2017.10.003. fogging. Sci. Hortic. 134, 210–221. https://doi.org/10.1016/J.
Tap, F., 2000. Economics-Based Optimal Control of Greenhouse Tomato Crop SCIENTA.2011.10.016.
Production. Wageningen University. Von Zabeltitz, C., 1999. Greenhouse structures. In: Stanhill, G., Enoch, H.Z. (Eds.),
Tap, F., Van Willigenburg, G., Van Straten, G., Van Henten, E.J., 1993. Optimal control of Ecosystems of the World 20: Greenhouse Ecosystems, 1st ed. Elsevier, pp. 17–69.
greenhouse climate: computation of the influence of fast and slow dynamics. IFAC Wallach, D., Martre, P., Liu, B., Asseng, S., Ewert, F., Thorburn, P.J., Van Ittersum, M.K.,
Proceedings Volumes 26 (2), 1147–1150. https://doi.org/10.1016/S1474-6670(17) Aggarwal, P.K., Ahmed, M., Basso, B., Biernath, C., Cammarano, D., Challinor, A.J.,
48650-2. De Sanctis, G., Dumont, B., Eyshi Rezaei, E., Fereres, E., Fitzgerald, G.J., Gao, Y.,
Tchamitchian, M., Van Willigenburg, G., Van Straten, G., 1992. Short term dynamic Zhang, Z., 2018. Multimodel ensembles improve predictions of crop-environment-
optimal control of the greenhouse climate. Wageningen MRS Report 92 (3). management interactions. Glob. Chang. Biol. 24 (11), 5072–5083. https://doi.org/
Thornley, J.H.M., France, J., 2007. Mathematical Models in Agriculture: Quantitative 10.1111/gcb.14411.
Methods for the Plant, Animal and Ecological Sciences, 2nd ed. CABI. Wallach, D., Makowski, D., Jones, J.W., Brun, F., 2019a. Model Evaluation. In: Working
Tiwari, G.N., 2003. Greenhouse Technology for Controlled Environment. Alpha Science. with Dynamic Crop Models, 3rd ed. Elsevier Academic Press, pp. 311–373. https://
Tiwari, G.N., Goyal, R.K., 1998. Greenhouse Technology: Fundamentals, Design, doi.org/10.1016/B978-0-12-811756-9.00009-5.
Modelling and Applications. Narosa publishing house. Wallach, D., Makowski, D., Jones, J.W., Brun, F., 2019b. Uncertainty and sensitivity
Udink ten Cate, A.J., 1983. Modelling and (Adaptive) Control of Greenhouse Climates. analysis. In: Working with Dynamic Crop Models, 3rd ed. Elsevier Academic Press,
Wageningen University. pp. 209–250. https://doi.org/10.1016/b978-0-12-811756-9.00006-x.
Vadiee, A., Martin, V., 2013. Energy analysis and thermoeconomic assessment of the Wasserstein, R.L., Lazar, N.A., 2016. The ASA statement on p -values: context, process,
closed greenhouse - the largest commercial solar building. Appl. Energy 102, and purpose. Am. Stat. 70 (2), 129–133. https://doi.org/10.1080/
1256–1266. https://doi.org/10.1016/j.apenergy.2012.06.051. 00031305.2016.1154108.
Van Bavel, C.H.M., Takakura, T., Bot, G.P.A., 1985. Global comparison of three Willmott, C.J., 1981. On the validation of models. Phys. Geogr. 2 (2), 184–194. https://
greenhouse climate models. Acta Horticulturae 174, 21–34. https://doi.org/ doi.org/10.1080/02723646.1981.10642213.
10.17660/actahortic.1985.174.1. Xu, D., Du, S., Van Willigenburg, G., 2018a. Adaptive two time-scale receding horizon
Van Beveren, P.J.M., Bontsema, J., Van Straten, G., Van Henten, E.J., 2015a. Minimal optimal control for greenhouse lettuce cultivation. Comput. Electron. Agric. 146,
heating and cooling in a modern rose greenhouse. Appl. Energy 137, 97–109. 93–103. https://doi.org/10.1016/j.compag.2018.02.001.
https://doi.org/10.1016/j.apenergy.2014.09.083. Xu, D., Du, S., Van Willigenburg, G., 2018b. Optimal control of Chinese solar greenhouse
Van Beveren, P.J.M., Bontsema, J., Van Straten, G., Van Henten, E.J., 2015b. Optimal cultivation. Biosyst. Eng. 171, 205–219. https://doi.org/10.1016/J.
control of greenhouse climate using minimal energy and grower defined bounds. BIOSYSTEMSENG.2018.05.002.
Appl. Energy 159, 509–519. https://doi.org/10.1016/j.apenergy.2015.09.012. Xu, D., Du, S., Van Willigenburg, G., 2019. Double closed-loop optimal control of
Van Henten, E.J., 1994. Greenhouse Climate Management: An Optimal Control greenhouse cultivation. Control. Eng. Pract. 85, 90–99. https://doi.org/10.1016/j.
Approach. Wageningen University. https://doi.org/10.1016/S0308-521X(94) conengprac.2019.01.010.
90280-1. Yang, J.M., Yang, J.Y., Liu, S., Hoogenboom, G., 2014. An evaluation of the statistical
Van Henten, E.J., 2003. Sensitivity analysis of an optimal control problem in greenhouse methods for testing the performance of crop models with observed data. Agric. Syst.
climate management. Biosyst. Eng. 85 (3), 355–364. https://doi.org/10.1016/ 127, 81–89. https://doi.org/10.1016/j.agsy.2014.01.008.
S1537-5110(03)00068-0. Zhang, G., Ding, X., Li, T., Pu, W., Lou, W., Hou, J., 2020. Dynamic energy balance model
of a glass greenhouse: an experimental validation and solar energy analysis. Energy
198, 117281. https://doi.org/10.1016/j.energy.2020.117281.
26