Professional Documents
Culture Documents
ScienceDirect
Research Paper
Ido Seginer a,*, Gerrit van Straten b, Peter J.M. van Beveren b,1
a
Technion, Haifa, Israel
b
Wageningen UR, Wageningen, The Netherlands
article info
Day-to-night heat storage in water tanks (buffers) is common practice in cold-climate
Article history: greenhouses, where gas is burned during the day for carbon dioxide enrichment. In Part
Received 24 October 2016 1 of this study, an optimal control approach was outlined for such a system, the basic idea
Received in revised form being that the virtual value (shadow price) of the stored heat (its ‘co-state’) could be used to
14 June 2017 guide the instantaneous control decisions. The results for daily-periodic weather showed: (1)
Accepted 22 June 2017 The optimal co-state is constant in time. (2) The optimal solution is associated with min-
Published online 24 July 2017 imum time on the storage bounds (buffer empty or full). With these conclusions as
guidelines, a semi-heuristic procedure of optimisation for realistic (i.e. not strictly periodic)
Keywords: weather is developed. The co-state remains constant while the storage trajectory is between
Greenhouse the heat storage bounds. It is gradually increased while the buffer is empty, and decreased
Heat buffer when the buffer is full, attempting to push the trajectory away from the bounds, thus
Optimal control minimising the time that the buffer is idle. The main outcomes are: (1) No information
Self-adjusting co-state about the future is required. (2) The algorithm changes the co-state automatically, pro-
CO2 enrichment ducing the correct annual variation (high in winter and low in summer). (3) The predictions
of yield and heat requirement compare favourably with practice. (4) The gain in perfor-
mance achievable with the suggested method is probably 75% or more of the true opti-
mum. (5) The procedure can be used in the design stage to determine the optimal buffer
size and the usefulness of other modifications of the system.
© 2017 IAgrE. Published by Elsevier Ltd. All rights reserved.
* Corresponding author.
E-mail address: segineri@tx.technion.ac.il (I. Seginer).
1
Now at B-Mex, Wageningen, The Netherlands.
http://dx.doi.org/10.1016/j.biosystemseng.2017.06.023
1537-5110/© 2017 IAgrE. Published by Elsevier Ltd. All rights reserved.
b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 1 8 8 e1 9 9 189
The main conclusions of the previous study, which relate increase the co-state gradually when storage is on the lower
directly to the current task, are: (1) The co-state of the optimal bound (empty buffer) and decrease it when on the upper
solution for periodic weather is constant. (2) Simulation-opti- bound (full buffer). The adjustment of the co-state at the
misation produces optimal solutions for periodic weather. (3) bounds is intended to push the trajectory away from the
The optimal co-state varies between seasons. (4) Performance bounds, thus minimising the time that the buffer is idle.
is improved by minimising the time that the heat buffer is In the following sections the system model is first briefly
completely empty or full. These results are now used to summarised, then applied to realistic year-long weather se-
develop a semi-heuristic approach to the more realistic quences to study some aspects of the control method.
problem of operation under actual weather. The general Finally, the effect of various parameters on the results is
argument is as follows: It has been shown that the co-state of explored.
the stored heat remains constant while the buffer is not
completely empty or full (storage not on bounds). It is also
clear that while the buffer is completely empty or full it is 2. Methods
ineffective in terms of operational storage. Hence a plausible
strategy might be to simulate-optimise with a constant co- The essentials of the system model and method of optimisa-
state while the storage trajectory is between the bounds, tion are described in the next five sub-sections. A fuller
190 b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 1 8 8 e1 9 9
description was provided in Seginer et al. (2017) but a sum- 2.2. Control variables
mary is included here for completeness.
Altogether there are 8 fluxes (Fig. 1) and 5 equality constraints
2.1. Heat and CO2 balances (Eqs. (2)e(6)), leaving 3 fluxes free as control (decision) vari-
ables. A sensible choice of the control variables is the venti-
A simplified schematic of the system is presented in Fig. 1. It lation rate, Q, and the heat fluxes HB and HS. Finding the
involves three compartments: Greenhouse (including crop), optimal control combination requires, therefore, a three
gas-fired boiler, and water filled heat-storage (buffer). The dimensional search over the feasible control domain. If the
only state variable to be considered is the stored heat, S. heat storage trajectory lies on either the upper or the lower
Heat fluxes are denoted by H and CO2 fluxes e by X. All bound (full or empty buffer), then
fluxes are one directional (positive), except HS, which is
positive when charging the buffer, and negative when dS=dt ¼ HS ¼ 0: (7)
discharging. This removes HS from the control vector, leaving just two
The balance equations of the compartments and of the 4- variables to search for.
way junction of Fig. 1 are (see Notation for symbols and units): The three control variables are constrained (limited) by
Buffer heat balance:
0 Q Qc ; (8)
dS=dt ¼ HS ; (1)
and
HG þ HF HT ¼ 0; (2)
HA ¼ ½absfHS g þ HS ð1 3 Þ=23 : (6) where G, N and Y are, respectively, the growth rates of the
whole plant, non-saleable material and saleable fruit. Y is
assumed to be proportional to G:
Y ¼ zG; (13)
h Li sCi
pfLi ; Ci g≡ LX ; (15)
hLX Li þ sCi
and
Fig. 1 e Schematic representation of a greenhouse system
with a gas fired boiler producing heat HB, and carbon qfTi g≡ 1 k Ti Tp 2 ; (16)
dioxide, XB. Some of the heat, HS, is transferred to a water-
where hLX , s, Tp and k are constants (see Notation). Respiration
filled heat-storage tank (buffer), some is lost from the
is formulated as
buffer, HA, some is recovered later (negative HS), and the
balance, HG, heats the greenhouse. HF is greenhouse solar RfTi g ¼ Rr expfbðTi Tr Þg; (17)
heating, HT is total greenhouse heat loss, XV is CO2 loss by
where Rr, Tr and b are constants. There are only two (tem-
ventilation, and G is (net) CO2 taken up by the crop. The
perature) constraints on the crop environment, namely
only state variable is the stored heat, S.
b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 1 8 8 e1 9 9 191
be zero (Eq. (7)), and the search is two dimensional, over Q and
Tmin Ti Tmax ; (18) HB. The control vector is then used to evaluate the greenhouse
indoor environment, as well as the increments of growth and
and there are no upper limits on light and CO2 concentration.
heat storage.
While 0 < S < Sc , the co-state, L, remains constant (Seginer
2.4. Greenhouse environment
et al., 2017, Eq. (34)), but as soon as S hits a bound, our heu-
ristic rule takes effect. The co-state is modified so as to push
The outdoor conditions are given in terms of hourly global
the storage trajectory away from the bound: When the tra-
(solar) radiation flux, Fo, hourly air temperature, To, and a
jectory is on the upper bound (full buffer), the co-state is
constant carbon dioxide concentration, Co. Outdoor photo-
slightly reduced each time step, to increase the incentive for
synthetic light, Lo, is derived from global radiation according
emptying the buffer, and vice versa when on the lower bound.
to
The adjustment rate of L, A (positive on the lower bound and
Lo ¼ hFL Fo ; (19) negative on the upper bound), is a parameter of the compu-
tation algorithm. All calculations in this study were done with
where hFL is a constant. The indoor light is obtained via this simulation-optimisation method, sometimes, however,
Li ¼ tLo ; (20) with A ¼ 0, namely with a strictly constant co-state. More
detail regarding the optimisation may be found in Seginer
where t is a constant. The formulation of HF, HT and XV is et al. (2017). In particular, the resolution of the search for the
optimal Hamiltonian was adapted from that study (Seginer
HF ¼ hFH Fo ; (21)
et al., 2017, Fig. 4).
Most of the calculations in this study were carried out for
ð1 þ BÞ
HT ¼ U þ rcðQ þ IÞ ðTi To Þ; (22) ‘realistic’ weather sequences: the so called SEL year, a syn-
B
thetic year-long hourly data-set characteristic of Dutch
weather (Breuer & Van de Braak, 1989), and actual weather
XV ¼ ðQ þ IÞðCi Co Þ; (23)
data of 10 years (2001e2010) at Hoek van Holland, The
where hFH, U, r, c, I and B are constants. While U, I and B are in Netherlands (Station 330; Royal Dutch Meteorological Insti-
reality variables, they are treated here as constants (B differs tute, KNMI). The results for the SEL weather were sometimes
between day and night). By properly substituting into Eqs. (2) compared with equivalent periodic sequences: (1) ‘natural’, a
and (3), explicit equations for Ti and Ci can be obtained. single-day periodic weather composed of hourly means across
the days of individual months, and (2) ‘pairs’, a two-day pe-
2.5. Performance criterion, Hamiltonian and riodic sequence where the first day is ‘natural’ multiplied by
computations 1.2, and the second day e by 0.8 (as in Seginer et al., 2017). For
each of the months the daily means of all approximations are
The (operational) performance criterion, J, of the greenhouse the same.
system is measured in terms of the value of the saleable fruit, The model parameter values have been collected from
minus the cost of control, over the relevant period of time various sources and are considered to be representative of
(here one month or one year). Capital and labour costs are not greenhouse tomato in The Netherlands. The complete set of
considered, assuming that they are independent of the control parameters defining the ‘standard’ case is given in Appendix A.
decisions (Seginer et al., 2017). Hence,
Z
J¼ ðuY Y uQ Q uB HB Þdt; (24) 3. Results
where uY, uQ and uB are, respectively, the unit value of the 3.1. Typical solution
saleable fruit and the unit costs of ventilation and of heat.
Note that the fruit-carbon, Y, is valued as soon as it is photo- Figure 2 shows the essentials of the solution of the standard
synthesised, meaning that green fruit is part of the ‘saleable’ case (Appendix A and SEL year with Co ¼ 0.015 mol [C] m3
fruit as defined by our model. The virtual value of the stored [air]). The top panel shows the co-state (continuous fluctu-
heat, the co-state, L, enters the decision process via the ating line) and the monthly potential gain from having a
Hamiltonian, H, which represents the incremental increase in buffer capacity of S ¼ 3 MJ [heat] m2 [ground] (circles). In
value of the greenhouse system as a whole (including stored winter the co-state fluctuates around the cost of boiler heat,
heat): uB ¼ 0.0085 $ MJ1 [heat], and in summer it fluctuates around a
small negative value. The monthly gain, G, has a maximum in
H ¼ LdS=dt þ uY Y uQ Q uB HB : (25)
March, a secondary maximum in September and is effectively
The control and storage trajectories are obtained by max- zero in mid-winter. These results are visually very similar to
imising the Hamiltonian at each time step (real or simulated) the results in Figs. 9 and 11 of Seginer et al. (2017) for periodic
over the feasible domain, with the current value of the co- weather based on the SEL year. The centre panel shows two
state L. When the stored heat is within its bounds, 0 < S < Sc aspects of the stored heat. The continuous line is the storage
(Eq. (11)), the (local) optimisation is attained by a three trajectory, S, fluctuating between the lower and upper bounds,
dimensional search over Q, HB and HS (Eqs. (8)e(10)). When the and the circles are the mean amount of heat stored (and later
buffer is empty or full (storage on the bounds), HS is known to recovered) per day in a given month. The stored amount
192 b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 1 8 8 e1 9 9
storage, HS þ HA , is lost directly to the atmosphere, HA, and shows that it should be larger for larger buffers. Starting with
how much is eventually dissipated via the greenhouse, HS. zero when the buffer is small, the optimal efficiency increases
The current model assumes a constant 3 throughout the steadily, until it hits the maximum, 1.0, when the capacity
whole year, but in principle 3 could be varied over time, for reaches about 4 MJ m2. As a result, the optimal gain, Curve 2,
instance by treating HA as an independent (fourth) control which by definition must be above or just touching the gain for
variable. Fig. 14 of Seginer et al. (2017) shows that the gain 3 ¼ 0.8, increases approximately linearly; the largest
from a buffer of a given size may be increased if 3 would be improvement over the standard case being when the buffer
high (~1.0) in winter and low (<0.5) in summer (positively vanishes ðSc ¼ 0Þ.
correlated with the co-state). This is intuitively reasonable, The improvement at Sc ¼ 0, almost 3.5 $ m2 y1, can be
because in summer the heat may become an undesirable by- explained as follows: It has been already established that most
product of CO2 enrichment, which, if possible, should be of the gain from having a buffer is related to CO2 enrichment
prevented from entering the greenhouse. Here, firstly the ef- in summer. Let us consider a summer day, a non-zero buffer
fect of buffer size on the economic (monetary) gain is and two extreme values of 3 : 0 and 1. If 3 ¼ 0, the buffer is
explored, then the interaction between size and storage effi- useless, but CO2 enrichment may still go on, while all the
ciency is investigated. generated heat, HB, either goes directly into the greenhouse
Figure 8 shows the effect of buffer size on the optimal per- via HG, or is wasted into the atmosphere via HA (Fig. 1). The
formance of a greenhouse system with a constant storage maximum gain from CO2 enrichment when 3 ¼ 0, is the
efficiency (here 3 ¼ 0.8). The main effects change asymptoti- intercept of Curve 2 in Fig. 9. If, on the other hand, 3 ¼ 1 and all
cally: the annual heat generation (gas consumption), HB, de- the daytime heat must go through the buffer, it very quickly
creases somewhat with increasing storage capacity, while the fills up and cannot accept more heat and hence stops
fraction stored ðHS =HB Þ increases from zero to almost half. As enrichment. If the daytime heat is delivered directly to the
the storage capacity, Sc, increases, the time the storage tra- greenhouse, the temperature in the greenhouse may become
jectory stays on the bounds diminishes to just a few percent of too high, unless a high ventilation rate is applied, which di-
the time, and the operational gain from having a buffer (not minishes the effect of CO2 enrichment. This should explain
considering capital cost), increases. The increasing gain Curve 4 of Fig. 9, where a large buffer may accumulate heat all
(mostly due to increased yield; not shown) is positively day to be used only at night when heat may be needed.
correlated with the amount of heat stored, HS (namely with Finally, if 3 could be controlled continuously, one would
CO2 concentration in the greenhouse), and negatively corre- expect the best possible operational outcome. Curve 3, based
lated with time on the bounds. on the best constant-per-month 3 (large in winter and small in
The gain of Fig. 8, for a constant storage efficiency, 3 ¼ 0.8 summer), is an approximation of that situation. The gain
(our standard), is copied into Fig. 9 as Curve 1 and compared represented by Curve 3 is, by definition, higher than that of
with results for more sophisticated storage systems. If any Curve 2, except when Sc ¼ 0, where 3 must be zero, as just
desired, yet constant, storage efficiency could be incorporated explained for Curve 2.
into a greenhouse system at the design stage, then Curve 4 The advantage of the variable-3 scheme over the constant-3
scheme is significant, provided that the additional investment
cost involved is not prohibitive. At this point in time we only
have an estimate of the investment cost of a standard
installation, which is represented in Fig. 9 by the sloping
dashed line at the top (~0.2 $ MJ1 y1; Vermeulen, 2012, p. 42). obvious that by doing so the (monetary) performance criterion
With this estimate of the installation cost, the optimal size of a J must decrease (expensive input, unchanged revenue), as is
standard system would be about Sc ¼ 4 MJ m2, somewhat shown in Fig. 11 (Criterion). It also results in a decreasing heat
larger than used here. The more sophisticated system with consumption and yield (not shown), indicating a reduction in
variable 3 would be more expensive, increasing the slope of cultivation intensity (reduced input and output). The useful-
the dashed line, and thus reducing the optimal size. If the ness of the buffer, however, increases as the heat becomes
optimal size of the variable-3 system is not smaller than more expensive (Fig. 11; Gain), presumably because although
Sc ¼ 2 MJ m2, it might still be better than a standard system of the amount stored diminishes as uB =uY increases, the value of
Sc ¼ 4 MJ m2. the stored heat increases (Fig. 10).
If all money-related variables were changed proportionally to Figure 12 compares the optimal monthly performance crite-
each other, then, obviously, the physical solution would rion of the SEL year with the mean monthly values for the
remain unchanged. The solution is only affected by a relative years 2001e2010 at Hoek van Holland. The standard de-
change in prices. In the present model, the three prices, uQ, uB viations about the monthly means are also shown. The mean
and uY and the adjustment rate, A, are money-related. How- is smoother than the SEL year, which on some of the months,
ever, as uQ is negligible (zero, Appendix A) and the effect of A, and most visibly in April, deviates from the mean by more
in the neighbourhood of the standard parameter set, is small than one standard deviation. Both curves are slightly skewed
(Fig. 4), the price-ratio heat/yield, uB =uY , is effectively the only to the left, compared to the global radiation curves, which are
factor which may influence the outcome of the control symmetrical around June (not shown). The annual criterion
calculations. for the SEL year is 44.7 $ m2 y1; 12% less than the mean
Changing the price ratio may affect decisions in both the criterion.
design and operation stages. The effect on the optimal size of
the buffer turns out to be negligible (not shown). The effect on
the optimal control decisions and the economic outcome is, 4. Discussion
however, very significant. Figure 10 shows four annual co-
state trajectories generated by our simulation-optimisation The discussion covers three items. Firstly, a summary list of
algorithm for three different levels of uB (0.010, 0.015 and 0.020 findings is presented. The capabilities of the proposed method
$ MJ1) and two levels of uY (0.5 and 1.0 $ mol1 [fruit-C]). are then briefly discussed, and finally it is placed in the general
Curves 1e3 (solid) are for uY ¼ 0.5 $ mol1 [fruit-C] and the context of co-state-based control.
dotted curve, 4, is for the same price ratio as Curve 1, except
that both uB and uY are twice as large. In all cases the winter 4.1. Summary of findings
value fluctuates around uB and the summer value ‒ around
zero. All curves are roughly proportional to one another. General properties of the co-state on-line adjustment method
While Case 4 produces the same annual sums for HB, HS and Y 1. The on-line, automatic method of adjusting the co-state,
as Case 1, and the criterion of Case 4 is twice as large as that of produces a useful sub-optimal solution to the day-to-
Case 1 (as expected), the corresponding co-states are quite night heat storage problem (Fig. 6).
different, mainly in winter.
The effect of prices on the co-state (Fig. 10) is not particu-
larly important. The question is whether, and how, the control
of the system is affected by changing the price ratio, for
instance by increasing the cost of heat, and thus of uB =uY . It is
Fig. 10 e On-line generated co-state trajectories for fruit Fig. 11 e Optimal performance criterion, J, and gain over a
price uY ¼ 0.5 $ mol¡1 [fruit-C] (standard) and three prices system without a buffer, G, as a function of the price ratio
of boiler generated heat: uB ¼ 0.010, 0.015 and 0.020 $ MJ¡1 heat/yield ðuB =uY Þ. The vertical dashed line is the standard
(continuous curves). The dotted curve is for uY ¼ 1.0 $ price ratio, 0.017 mol [C] MJ (the crossing of all three curves
mol¡1 [fruit-C] and uB ¼ 0.020 $ MJ¡1 SEL year. at a point is just a coincidence). SEL year.
b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 1 8 8 e1 9 9 197
Right now, the method as suggested has several obvious these requirements. By changing the values of certain co-
strengths: states, a grower could correct (compensate) on-line for inac-
curate parameters of the underlying model, for an unexpected
1. It is simple to implement on-line. weather spell, or for changing market conditions. The main
2. The adjustment of the co-state, L, to the varying weather, reason that changing the co-states can compensate for model
and in particular to the change of seasons, is automatic. inaccuracies (e.g., inaccurate 3 ) is that an erroneous model
3. The strategy does not require any information about the often just affects the rate of progress (here storage rate), which
future. The co-state is adjusted according to the current can be usually rectified by changing the appropriate co-state
weather, prices, state, S, and co-state, L, the latter two value (Seginer, 2013). Eventually (at the end of the season),
depending on the recent history of the system. the model may be re-calibrated.
If the heat buffer is consistently full (or empty), reducing
If a reliable system model is lacking, measured heat actually (or increasing) the incentive to ‘fill’ should rectify the situa-
stored in the buffer could be used to adapt the co-state on-line, tion. The dial (co-state) adjustment could be manual or auto-
similarly to Seginer (2013). matic. In the heat storage problem, where automatic
monitoring (feedback from buffer) is simple (unlike esti-
4.3. Greenhouse control by co-state adjustment mating, for example, the ‘vigour’ of a crop), the adjustment
could be automatic.
The day-to-night heat storage is but one of many greenhouse Co-states are generally better agents of control policy
process-control problems. It differs from seasonal growth transfer from one situation to the other than are control fluxes
problems by focussing on the short term scale and having or set-points (Van Straten, van Willigenburg, van Henten, &
state bounds (constraints) which are reached almost every van Ootegem, 2011, Section 9.4.7.4). The co-states are often
day. Otherwise, however, when observed from the co-state constant, night and day (as in our case), unlike the set-points
point of view, they all have much in common, and in partic- and control fluxes, and remarkably robust with respect to the
ular the distinction between the strategic and tactical levels, year-to-year variation in weather (Van Henten, 1994; our re-
which will now be briefly described. sults for 2001 to 2010, not shown). Furthermore, a single co-
Conventional computerised control of greenhouses (Kamp state may generate several control fluxes (here three: Q, HB
& Timmerman, 1996), requires the setting of numerous pa- and HS). Finally, having an economic connotation, such as
rameters, which, in turn, define the set-points for the various LyuB in winter and Ly0 in summer, is also helpful.
indoor environmental variables, such as temperature.
Adjusting these parameters manually is problematic, mainly
because the interactions between their effects are often rather
Acknowledgement
complex. At the same time, growers are reluctant to relin-
quish all control to the computer, partly because the computer
We thank Ep Heuvelink of WUR for important information
models have proved not to be all that reliable, and partly
regarding the tomato crop.
because the weather, biology, or economics may turn out to be
different than initially anticipated. A typical grower makes
Appendix A. Parameter values for the standard
set-point adjustments, when deemed necessary, based mostly
case
on observations of the state of the crop (size, developmental
stage, vigour, infestation, etc.) and possibly other information
The following are the standard parameter values. The stan-
(prices, logistics, weather, etc.). This information is mentally
dard case utilises the SEL year weather data. When non-
translated into processes (growth, development, transpiration,
standard values are used, this is indicated in the text and
infection etc.) to be accelerated or retarded by changing the
figures. The parameters are arranged alphabetically (Latin,
set-points. The actions taken by the grower are usually meant
then Greek) within each topic.
to affect the long-term development of the crop, leaving the
short term adjustments to the computer. This means that the
Crop growth (Eqs. (13), (15)e(18))
grower makes the strategic decisions, while the computer is
f ¼ 0.91
responsible for the tactical level. With this in mind, the
Rr ¼ 0.75e6, mol [C] m2 [ground] s1
problem of adjusting a large number of set-points may,
Tmin ¼ 12, C
perhaps, be replaced by adjusting a limited number of dials,
Tmax ¼ 33, C
each associated with a certain process, which the grower can
Tp ¼ 23, C
turn up or down, indicating which process to accelerate and
Tr ¼ 0, C
which to retard. These dials may be labelled ‘growth’, ‘devel-
b ¼ 0.0693, K1
opment’, or ‘botrytis’, depending on relevance to a particular
hLX ¼ 0.07, mol [C] mol1 [PAR]
situation. In the present context it could be a ‘fill (buffer)’ dial.
z ¼ 0.7
It seems that the co-states are the equivalents of such
k ¼ 0.0015, K2
dials. The proper choice of the few most appropriate state
s ¼ 0.004, m s1
variables, and hence the associated co-states, is critical. They
should be relevant to the problem, meaningful to the grower,
Greenhouse environment (Eqs. (19e22))
observable (measurable) and controllable (respond to the avail-
Bday ¼ 0.3
able controls). In our case, the heat stored in the buffer meets
b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 1 8 8 e1 9 9 199