Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

Energy Conversion and Management 64 (2012) 473–481

Contents lists available at SciVerse ScienceDirect

Energy Conversion and Management


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

Modeling of existing cooling towers in ASPEN PLUS using an equilibrium


stage method
João A. Queiroz a,b,c, Vitor M.S. Rodrigues b, Henrique A. Matos c, F.G. Martins a,⇑
a
LEPAE, FEUP-DEQ, Universidade do Porto, R. Dr. Roberto Frias, 4200-465 Porto, Portugal
b
Dow Portugal, Produtos Químicos, S.U.L., R. do Rio Antuã, n.1, 3860-529 Estarreja, Portugal
c
CPQ, IST-DEQ, Universidade Técnica de Lisboa, Av. Rovisco Pais, 1049-001 Lisboa, Portugal

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

Article history: Simulation of cooling tower performance considering operating conditions away from design is typically
Received 12 January 2012 based on the geometrical parameters provided by the cooling tower vendor, which are often unavailable
Received in revised form 29 March 2012 or outdated. In this paper a different approach for cooling tower modeling based on equilibrium stages
Accepted 29 March 2012
and Murphree efficiencies to describe heat and mass transfer is presented. This approach is validated
Available online 1 October 2012
with published data and with data collected from an industrial application. Cooling tower performance
is simulated using ASPEN PLUS. Murphree stage efficiency values for the process simulator model were
Keywords:
optimized by minimizing the squared difference between the experimental and calculated data using the
Cooling tower
ASPEN PLUS
Levenberg–Marquardt method. The minimization algorithm was implemented in Microsoft Excel with
Simulation Visual Basic for Applications, integrated with the process simulator (ASPEN PLUS) using Aspen Simulation
Equilibrium stages Workbook. The simulated cooling tower air and water outlet temperatures are in good accordance with
Murphree efficiency experimental data when applying only the outlet water temperature to calibrate the model. The method-
ology is accurate for simulating cooling towers at different operational conditions.
Ó 2012 Elsevier Ltd. All rights reserved.

1. Introduction the reduction of water flow rate by evaporation was neglected in


the energy balance.
Cooling towers are widely employed in many industrial applica- Kloppers and Kröger [3] evaluated three methods used in cool-
tions for rejecting waste heat from the process to the environment. ing tower design, namely, Merkel, Poppe and effectiveness-NTU
The principle behind a cooling tower operation is evaporative cool- and gave a detailed derivation of the heat and mass-transfer equa-
ing which, in theory, would allow circulating water to equal ambi- tions of evaporative cooling in cooling towers. Based on Merkel
ent air wet-bulb temperature. Evaporative cooling is a process with equation, Picardo and Variyar [4] presented a power law that re-
simultaneous mass and heat transfer between air and circulating lated packed height with excess air and determined equation
water. parameters for air wet-bulb temperature between 10 and 34 °C
There are several methods and strategies related to the model- and cooling range between 40 and 20 °C. They also showed that be-
ing of cooling towers with different levels of complexity. According yond a certain air flow the reduction in packed height does not jus-
to Jin et al. [1], the first theoretical analysis of cooling towers was tify the increase in energy utilization for air compression.
performed by Dr. Fredrick Merkel in 1925. He proposed a theory Castro et al. [5] developed an optimization model for a cooling
relating evaporation and sensible heat transfer where there is water system composed of a counter flow tower and five heat
counter flow contact of water and air. As described by Benton exchangers where the thermal and hydraulic interactions in the
et al. [2] Merkel expressed the number of transfer units (NTU) as overall process were considered. They observed that forced with-
a function of the integral of the water temperature difference di- drawal of water upstream of the tower is an important resource
vided by the enthalpy gradient where, to reduce the governing for fulfilling cooling duty requirements. Khan et al. [6] presented
relationships to a single separable ordinary differential equation, a fouling growth model where it was demonstrated that the effec-
several simplifying assumptions were made: Merkel assumed that tiveness of a cooling tower degrades significantly with time, indi-
the Lewis factor, relating heat and mass transfer was equal to 1; cating that for a low fouling risk level (p = 0.01), which is the
the air exiting the tower was saturated with water vapor; and probability of fill surface being fouled up to a critical level after
which a cleaning is needed, there is about 6.0% decrease in effec-
tiveness. Al-Waked and Behnia [7] applied computational fluid
⇑ Corresponding author. Tel.: +351 22 508 1974; fax: +351 22 508 1449. dynamics (CFD) for natural draft wet cooling tower. The difference
E-mail address: fgm@fe.up.pt (F.G. Martins). between outlet air temperature predicted by the CFD model and

0196-8904/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.enconman.2012.03.030
474 J.A. Queiroz et al. / Energy Conversion and Management 64 (2012) 473–481

Nomenclature

F(k) sum of squared error between experimental and model e allowable difference between two consecutive itera-
prediction values tions
G humid air mass flow (kg/h) k model parameters
H[F(k)] Hessian matrix of F(k) rF(k) gradient of F(k)
I identity matrix
L water mass flow (kg/h) Subscripts
m number of data points in each sub-set i relative to component i
N number of equilibrium stages j relative to stage j
P pressure (kPa) k kth iteration
RH relative humidity (%) l lth experimental point
s search direction test relative to test data sub-set
T temperature (°C) trn relative to train data sub-set
x independent variables val relative to validation data sub-set
y vapor composition on equilibrium stage
Y model output Superscripts
Y average value of the model output atm relative to ambient conditions
b
Y experimental value a relative to air stream
b control coefficient in Levenberg–Marquardt method exp experimental data
Dk increment of k mod model predicted values
w relative to water stream

design results was less than 3%. Jin et al. [1] proposed a model 2. Problem statement
based on heat resistance and energy balance principles where
empirical parameters were introduced, avoiding the need to spec- When simulating a cooling water network with a process simu-
ify geometrical parameters. Rubio-Castro et al. [8] determined lator such as ASPEN PLUS there is no standard procedure for setting
optimal cooling tower design parameters and temperature profiles up the cooling tower block. This means that when process alterna-
across a counter flow cooling tower by applying a rigorous heat tives are being studied or real-time optimization is being per-
and mass transfer model. Non-linear algebraic equations were formed, there is no direct way of evaluating the impact that any
solved using a discretization approach with a fourth-order Run- change in the cooling water temperature and flow or change in
ge–Kutta algorithm. Given a set of experimental data to train the ambient air conditions has on the cooling tower performance
model, Hosoz et al. [9] suggested that applying artificial neural net- and, consequently, on the process itself.
works (ANNs) for modeling the cooling tower performance avoided The aim of this work is therefore to describe a methodology that
the solution of complex differential equations. Predicted and enables the simulation of a counter flow, induced draft, cooling
experimental values had correlation coefficients in the range of tower in ASPEN PLUS. The outcome of the proposed approach is a
0.975–0.994 and mean relative errors in the range of 0.89–4.64%. model that simulates the behavior of a real cooling tower and is
Pan et al. [10] presented a data-driven model-based assessment capable of working as a stand-alone model or be integrated into
strategy to investigate the performance of an industrial cooling a larger simulation model. It should provide grounds for optimiza-
tower. Considering 1 month test interval and based on water mass tion studies, where off-design conditions such as water and air
flow rate, water inlet temperature, air dry-bulb temperature, rela- flow variations can be simulated; debottlenecking studies, where
tive humidity and fan motor power consumption the predicted equipment limitations may be of interest as production rates are
water outlet temperature was within a ±5% error band and pre- increased; and operability studies, where evolution of equipment
sented a mean square error of 0.29 °C. Serna-González et al. [11] performance can be assessed to help plant personnel troubleshoot
used mixed-integer non-linear programming (MINLP) techniques operations.
to evaluate the optimal conditions of a mechanical draft cooling Given a set of data consisting of water and air inlet tempera-
tower that minimize the total annual cost for a given heat load, ture, water and air inlet flow, air inlet humidity and ambient
dry- and wet-bulb inlet air temperatures and temperature con- pressure the problem then consists in determining model param-
straints on the cooling water network. Rao and Patel [12] com- eters that mimics actual cooling tower performance. Although it
pared the results obtained by Serna-González et al. [11] with the is possible to use the model to calculate evaporation ratio and
ones achieved when applying an artificial bee colony algorithm. cooling tower heat duty, only the output variables which could
Using the artificial bee colony algorithm resulted in an objective be compared with the published data by Simpson and Sherwood
function value lower than the one achieved by Serna-González [13] i.e., water and air outlet temperatures, were presented.
et al. [11] for all six case studies (improvement between 1.27% Moreover, the model was developed considering the following
and 11.17%). assumptions:
As an alternative to the abovementioned methodologies a dif-
ferent approach that does not involve the solution of differential 1. Pressure drop across the cooling tower was not considered
equations can be used to model a cooling tower operation by because the model was not used for design purposes and it
applying an equilibrium stage. While the equilibrium stage method has minor effect when compared to ambient pressure.
can hardly be used for design purposes without proper correlations 2. The operation of cooling tower was assumed as an adia-
that allow the determination of the height equivalent to a theoret- batic process.
ical plate (HETP) it is demonstrated in this work that both the out- 3. The water stream was considered to be pure as impurities
let water and air temperature predicted by the model are quite in this stream do not significantly change equilibrium
accurate when compared to the experimental values. properties.
J.A. Queiroz et al. / Energy Conversion and Management 64 (2012) 473–481 475

4. The cooling tower was in steady-state operation. Air out (T a out, Gout, RHout)

3. Modeling of a cooling tower in ASPEN PLUS


Water in (T w in, Lin)
The proposed cooling tower model is implemented in ASPEN
PLUS [14]. This process simulator provides several built-in model
blocks that can be directly applied in process simulation. Addition- Stage N-1
ally, this process simulator has an extensive physical property
database where the stream properties required to model the mate-
Stage N
rial streams in a plant are available.
Air in (P atm, T a in, G in, RH in)
For simulating the cooling tower in ASPEN PLUS an approach
based on equilibrium stages is applied. For this purpose ASPEN
PLUS provides the RADFRAC built-in block considering neither
reboiler nor condenser, which allows the calculation of the liquid Water out (T w out, Lout)
and vapor/gas equilibrium on each equilibrium stage.
Fig. 1. Representation of a cooling tower with N equilibrium stages. The air stream
3.1. Step 1 – set up the model in ASPEN PLUS enters the cooling tower at the bottom stage and leaves at the top stage; it is
characterized by its dry bulb temperature (Ta), flow (G) and relative humidity (RH).
Water stream enters the cooling tower at the top stage and leaves at the bottom
Cooling towers operate at relatively mild temperature and pres-
stage; it is characterized by its temperature (Tw) and flow (L).
sure so it can be assumed that the behavior of the thermodynamic
equilibrium between water and air is close to ideality. This fact is
confirmed comparing the results achieved using the cooling tower parameters for a given cooling tower by calculating the number of
general model when applying the ideal property method and other equilibrium stages and Murphree efficiency of each stage. Heat ex-
method that takes into account non-idealities (NRTL). Moreover, changer performance strongly depends on the cooling medium sup-
considering a base case data and two scenarios with ±10% varia- ply temperature, therefore the most relevant parameter when
tions on RHin, Ta in, Tw in, Lin and Gin the results are identical inde- considering a cooling tower operation is the water temperature
pendently of the chosen property method (Table 1). Thus, it can (Tw out) that this equipment is able to provide given certain opera-
be concluded that assuming an ideal behavior of the system sub- tional conditions. Model parameters values are adjusted by using
stances (water and air) is a good approximation for this process. a set of experimental data (l = 1, . . . , m) and by minimizing objective
For defining the RADFRAC block in ASPEN PLUS the following function (F), which is the sum of the squared difference between
parameters were specified in ASPEN PLUS: experimental, T w l
out;exp
, and model outputs, T w
l
out;mod
(Eq. (2)).
X
m
out;mod 2
– Calculation type set as ‘Equilibrium’. F¼ ½T w
l
out;exp
 Tw
l  ð2Þ
– Inlet water entered the block ‘Above-Stage’ on the first stage. l¼1
– Outlet water left the block on the last stage.
Eq. (2) can be rewritten as Eq. (3) to give a more general formulation
– Inlet air entered the block ‘On-stage’ on the last stage.
of the function to be minimized [15]. Each experimental measure-
– Outlet air left the block on the first stage.
ment is described by xl, which is the independent variables vector
– Reboiler and condenser set as ‘None’.
and corresponds to the measured inlet streams values (Table 2).
– Efficiency type set as ‘Murphree efficiencies’ on each stage.
Model output, T lw out;mod (xl, kk), specified in Table 2 depends on both
the independent variables, xl, and the model parameters, kk. The
Murphree efficiency is applied when accounting for deviations
minimization of the objective function is accomplished by applying
from ideality, i.e., considering that in an equilibrium stage, liquid
the method described in this work, which consists on an iterative ap-
and vapor phases do not reach thermodynamic equilibrium. Eq. (1)
M proach thus; model parameters are adjusted for each kth iteration.
defines Murphree vapor efficiency, Eff i;j , for component i on stage j,
Ideally the model would return an output, T w l
out;mod
ðxl ; kk Þ, equal
where yi,j and xi,j are the vapor and liquid composition, respectively,
to the measured value, T w l
out;exp
, given a certain set of inlet param-
and yi;j is the composition of the vapor that would be in equilibrium
eters, xl. Therefore, for m experimental data points, the goal is to
with the liquid leaving the equilibrium stage. Fig. 1 shows the
find the model parameters that correspond to the minimum value
schematic representation of an equilibrium stage in a cooling tower.
of the objective function described by Eq. (3), F(kk).
M yi;j  yi;jþ1
Eff i;j ¼ ð1Þ X
m
yi;j  yi;jþ1 Fðkk Þ ¼ ½T w out;exp
 T w out;mod ðxl ; kk Þ2 ð3Þ
l
l¼1

3.2. Step 2 – determination of model parameters Regarding the particular case of a cooling tower the parameter
vector (kk) to be adjusted is the number of equilibrium stages, N,
M
Once the general model representing the cooling tower has been and Murphree stage efficiencies, Eff i;j . Hence, the minimization
implemented on ASPEN PLUS it is necessary to determine the model problem is subject to the following constraints:

Table 1
Comparison between the outputs of the model considering NRTL and IDEAL property methods. The number of equilibrium stages was set to 2 and Murphree stage efficiencies to
1. Atmospheric pressure was kept constant at 101.3 kPa.

Inputs Outputs
RHin (%) Ta in
(°C) Tw in
(°C) Lin (kg/h) Gin (kg/h) Tw out
(°C) Ta out
(°C)
NRTL IDEAL NRTL IDEAL
Base case 80 30 37 7.5 10 29.5 29.5 32.4 32.4
+10% 88 33 41 8.3 11 33.3 33.3 35.9 35.9
10% 71 27 33 6.7 9 25.6 25.6 28.6 28.6
476 J.A. Queiroz et al. / Energy Conversion and Management 64 (2012) 473–481

Table 2 of stages should be incremented in order to avoid an inconsistent


List of independent, dependent and model variables. stage efficiency value. The algorithm must then be reinitialized,
Independent variables [x] Model variables [k] Dependent variables [Y] taking into account the new number of equilibrium stages (N + 1).
– Inlet air temperature – Stage 1 efficiency – Outlet water
M
temperature 3.2.2. Step 2.2 – Murphree stage efficiencies (Eff w;j )
– Inlet water – Stage 2 efficiency
Levenberg–Marquardt is in nature an improved Gauss–Newton
temperature
– Inlet air flow ..a method by incorporating steepest-descent method into the
. iterative update scheme, using a search direction between these
– Inlet water flow – Stage N efficiency
two methods. In the Levenberg–Marquardt method the search
– Inlet air humidity
– Atmospheric pressure direction, sk, is determined by solving Eq. (4), where H[F(kk)] is
a
the Hessian matrix of F(kk), bk is the control coefficient of the
The number of model variables depends on the number of equilibrium stages.
Levenberg–Marquardt method, I is the identity matrix and rF(kk)
is the gradient of F(kk).
M When bk tends to zero Levenberg–Marquardt method ap-
s:t: 0 < Eff i;j 6 1 proaches the Gauss–Newton method whereas when bk tends to
N>1 infinity the Levenberg–Marquardt method approaches the steep-
N is integer est-descent method. The values of bk during the iterative process
are chosen in the following way: when initializing the algorithm
To determine model parameters an algorithm was established, bk is set to a large value so that the Levenberg–Marquardt method
as shown in Fig. 2. This algorithm was implemented in Microsoft manifests the robustness of the steepest-descent method, meaning
Excel, using Visual Basic for Applications, and simulations were that the initial guess can be chosen with less caution. For each iter-
performed in ASPEN PLUS. Information flow between ASPEN PLUS ation, if F(kk + sk) < F(kk), convergence is accelerated by decreasing
and Microsoft Excel was enabled by Aspen Simulation Workbook. bk by a certain amount set by the user; otherwise, bk is increased
Considering the set of experimental data regarding the opera- in order to enlarge the searching area [16].
tion of a given cooling tower the first step is to divide it into three
sub-sets: training (trn), validation (val) and test (test). Training and ðH½Fðkk Þ þ bk IÞsk ¼ rFðkk Þ ð4Þ
validation sub-sets are used to determine model parameters and The objective function is not described by an analytical expres-
are inputs of the algorithm; the test sub-set is used to confirm sion; therefore, the gradient is calculated using the numeric meth-
whether the model parameters returned by the algorithm are ade- od given by Eq. (5). The step size, Dk, is an infinitesimal positive
quate when applied to an independent set of data. number that enables the numerical calculation of the gradient.
When considering only a training sub-set to adjust model
parameters, each iteration would bring the difference between ðFðkk þ Dkk Þ  Fðkk  Dkk ÞÞ
rFðkk Þ ¼ ð5Þ
model and experimental values closer to zero. However, the incon- 2  Dkk
venient of this approach is that there would be a point where the
Regarding the Hessian matrix, the calculation is performed by
model is too adjusted to that specific set of data, resulting in a
applying the expanded Taylor series truncated to the second term
model very dependent of the data set used to train the model. To
as shown in Eq. (6).
avoid this situation a second set of data, the validation sub-set
(val), is used to avoid the model overfitting. During the starting 1 T
Fðkk þ Dkk Þ ¼ Fðkk Þ þ rFðkk Þ  Dkk þ  k  H½Fðkk Þ  Dkk ð6Þ
iterations it is expected that the error between model and experi- 2 k
mental values decreases for both training and validation sub-sets
as model parameters start to be adjusted. Despite the fact that
the objective function value for the training sub-set decreases as 3.3. Step 3 –ASPEN PLUS and Microsoft Excel interaction
the algorithm progresses (Ftrn(kk+1) < Ftrn(kk)), there can be point
where the value of the objective function regarding the validation As stated in Steps 1 and 2 the model outputs – which corre-
sub-set reaches a local minimum and then starts to increase. Be- spond to the outlet water temperature – are generated running
yond this point it is considered that the model becomes dependent an ASPEN PLUS simulation model. These results generated by the
of the training sub-set data and the inner algorithm stops (Fig. 2), process simulator are fed into the minimization algorithm imple-
thus assuming that the model parameters corresponding to this mented in Excel, which in turn will feed the simulator with new
minimum is the most adequate. model parameters (Fig. 3).
This two-way connection between the process simulator (AS-
PEN PLUS) and the minimization algorithm (Visual Basic for Appli-
3.2.1. Step 2.1 – number of equilibrium stages (N) cation in Microsoft Excel) is provided by an interface between both
The problem, as described, represents a mixed integer non-lin- software tools. The interface is guaranteed by Aspen Simulation
ear programming (MINLP) problem. The algorithm introduced in Workbook [17] which allows a seamless data transfer between AS-
this work transform the MINLP into a non-linear problem (NLP) PEN PLUS and Microsoft Excel.
model by setting a fixed value to the number of stages, eliminating Aspen Simulation Workbook is a tool for interfacing Aspen-
the integer constraint. Fig. 2 shows how the algorithm is structured Tech’s process simulation models with Microsoft Excel worksheets.
and how the number of equilibrium stages initial estimate is con- Aspen Simulation Workbook also has tools to link model variables
firmed at the end of inner algorithm. to plant data tags imported using third-party applications. These
With this approach, the NLP minimization problem can be capabilities allow modeling experts to link models and plant data
solved by using the Levenberg–Marquardt search method. The ini- and publish the resulting models as Excel worksheets [18].
tial guess for the number of equilibrium stages is set to be the min-
imum allowed by ASPEN PLUS RADFRAC block, N = 2. Murphree 3.4. Step 4 – implementation of the proposed methodology
efficiency for each stage is then determined using the Levenberg–
Marquardt method and, if the algorithm generates a result corre- A step by step flowchart representing the actions that must be
sponding to Murphree stage efficiencies higher than 1, the number taken to implement the proposed methodology is shown in Fig. 3.
J.A. Queiroz et al. / Energy Conversion and Management 64 (2012) 473–481 477

Fig. 2. Algorithm for the determination of model parameters. The inner circle of the algorithm is a NLP problem whereas if the number of equilibrium stages is considered it is
a MINLP problem.

Due to the comprehensive experimental data as well as detailed A third set of data (Table 3) was used to confirm the applicabil-
information about the experimental setup the work of Simpson ity of the proposed approach to a real industrial application, which
and Sherwood [13] is often used to evaluate the appropriateness corresponded to a set of experimental data collected from an
of cooling tower models [6,19,20]. These authors published exper- industrial mechanical induced draft cooling tower. The industrial
imental data regarding the operation of two mechanical induced cooling tower with the characteristics specified in Table 4 belongs
draft cooling towers, designated by tower R-1 and tower R-2. to one of the manufacturing plants of Dow Chemical Company in
Although the ambient pressure is not mentioned in the work of Portugal. Water pumps and fan motors are equipped with fixed
Simpson and Sherwood [18], taking into account the nature of speed drives so that the water and air flow rate are kept constant.
the work it is assumed that it remained approximately constant The dry-bulb temperature and relative humidity were measured
during the length of the experimental work. These two sets of pub- near the air entrance of the cooling tower with a data logger
lished experimental data are used for validating the approach pro- (Tinytag View 2 – TV4500). In- and outlet water temperature are
posed in this work. both registered by an online process monitoring system in the
478 J.A. Queiroz et al. / Energy Conversion and Management 64 (2012) 473–481

Table 3
Collect plant data
Experimental and modeled values for Dow’s cooling tower collected during 1 month
(make sure equipment is operating in
steady state conditions) period (20 May 2011–20 June 2011). Air and water flow correspond to design. Air
outlet temperature was not monitored.

Patm (kPa) Ta in,exp RHin Tw in,exp Tw out,exp Tw out,mod


Build ASPEN PLUS simulation Step 1 (°C) (%) (°C) (°C) (°C)
model
Train
101.44 17.45 97.00 28.46 22.87 22.93
101.95 15.97 91.65 27.38 21.74 21.63
Implement minimization Step 2 101.92 21.29 72.80 29.91 23.82 23.95
algorithm in Microsoft Excel 101.85 20.26 67.62 28.79 22.70 22.85
102.12 17.03 87.15 28.73 22.50 22.47
101.11 21.22 85.47 31.01 25.14 25.02
Link both tools with Aspen 100.63 18.09 89.67 28.61 22.75 22.93
Step 3
Simulation Workbook 102.19 19.71 79.10 29.77 23.55 21.83
102.12 21.98 73.90 30.86 24.51 24.69
102.09 17.41 80.22 28.73 22.36 22.37
101.51 19.48 88.00 28.07 23.03 23.16
Run algorithm for determining 101.14 19.71 79.10 29.77 23.55 23.58
model parameters 101.82 18.65 85.77 27.76 22.61 22.61
101.78 17.79 92.47 29.31 23.30 23.23
101.92 22.02 68.17 30.49 24.20 24.26
102.02 19.36 81.87 30.02 23.68 23.73
Model parameters 100.90 17.60 97.85 29.83 23.72 23.59
Experimental inputs 101.14 17.62 81.60 27.65 21.83 21.97
101.78 21.15 65.45 29.87 23.17 23.51
101.58 27.27 37.98 30.88 24.49 24.37
Aspen Simulation Workbook

101.82 18.65 85.77 27.76 22.61 22.62


ASPEN PLUS Microsoft Excel 101.78 26.93 35.40 30.31 24.06 23.82
101.41 19.72 78.55 27.29 22.40 22.47
Process simulator model Minimization algorithm
101.65 19.50 68.97 28.81 22.74 22.61
101.31 25.47 41.88 30.63 24.29 23.94
101.44 22.65 74.97 31.15 24.86 25.14
Experimental Inputs

Experimental Outputs

Model Outputs

100.90 16.46 79.10 28.15 21.92 21.65


Model parameters

101.51 24.76 51.58 30.91 24.65 24.44


101.75 25.43 65.72 31.40 25.64 25.82
101.78 21.75 75.52 28.93 23.65 23.89
Validation
102.02 21.85 62.77 30.11 23.78 23.75
101.82 17.40 86.60 29.14 22.80 22.79
101.88 19.79 62.22 28.95 22.64 22.49
102.02 21.85 67.35 30.32 23.97 24.50
Model outputs
101.31 22.30 75.25 31.39 25.26 25.12
102.12 17.61 83.25 29.17 22.80 22.77
102.02 21.89 73.35 30.82 24.44 24.62
Fig. 3. Overall procedure flowchart showing the interaction between the process
100.87 17.55 96.72 29.50 23.51 23.40
simulator (ASPEN PLUS) and the minimization algorithm (implemented in Micro-
101.88 19.79 62.22 28.95 22.64 22.49
soft Excel), being this interaction mediated by Aspen Simulation Workbook. The
101.99 22.16 72.25 30.57 24.51 24.56
steps referred in this figure are described in detail in Section 3 of this work.
Test
101.88 19.79 62.22 28.95 22.64 22.51
101.78 22.17 73.90 30.32 24.51 24.53
main headers. Since the atmospheric pressure is not monitored on 101.85 18.07 85.20 29.44 23.16 23.12
site this variable was retrieved from two weather stations 20 km 101.82 21.31 76.90 28.14 23.23 23.40
away on opposite directions and the mean value was computed. 101.78 20.71 75.25 30.17 24.05 23.97
The atmospheric pressure was obtained from Weather Under- 101.75 22.60 52.35 27.85 22.53 22.43
101.78 20.71 52.35 28.11 22.03 21.91
ground website [21]. Due to the fact that the air outlet temperature
101.99 22.57 65.72 30.90 24.42 24.49
is not monitored the objective function to be minimized only takes 102.02 16.31 83.55 28.10 21.79 21.81
into account the outlet water temperature. 101.07 22.44 78.55 31.04 25.04 25.16
To train, validate and test the model, experimental data was di-
vided in three sets. Data points were randomly split for training
(trn), validating (val) and testing (test) with 60/20/20%, respec- Table 4
tively. Data sets regarding tower R-1, R-2 and Dow’s were defined Dow’s cooling tower design specifications.
by 49, 45 and 50 experimental measurements, respectively. Type Induced draft counter flow
As stated in the algorithm shown in Fig. 2, the procedure stops
Tower dimensions
when the error associated to the validation sub-set increases or, for H 12.85 m
the case when the this value is continuously decreasing, the differ- W 13.32 m
ence between the error associated to the training sub-set of two L 37.58 m
consecutive iterations is less than a pre-set value (e). The stopping Packing height 5.1 m
Design water flow 5678 tonne/h
criteria that was verified for both tower R-1 and Dow’s tower was Design water inlet temperature 36.7 °C
Fval(kk+1) > Fval(kk) and for tower R-2 Ftrn(kk+1)  Ftrn(kk) < e. Figs. 4–6 Design water outlet temperature 30.0 °C
show the evolution of the objective function in regard to the num- Design air inlet wet-bulb temperature 26.7 °C
ber of iterations. Design air flow 4505 tonne/h
Number of fans 3
Considering the algorithm for model identification described
Nominal fan power (each) 75 kW
before it is stated that additional equilibrium stages are added only
J.A. Queiroz et al. / Energy Conversion and Management 64 (2012) 473–481 479

Fig. 4. Objective function evolution for tower R-1 for training [j] and validation Fig. 6. Objective function evolution for Dow’s tower for training [j] and validation
[h] sub-sets. [h] sub-sets.

Murphree stage efficiencies for the three cooling towers were:


M M M M
Eff w;1 = 0.93, Eff w;2 = 0.85 for tower R-1; Eff w;1 = 0.89, Eff w;2 = 0.74
M M
for tower R-2 and = 0.96, Eff w;1 Eff w;2
= 0.70 for Dow’s tower.
The good agreement between predicted and experimental can
verified in Figs. 7–9 for water outlet temperatures.
Due to the fact that air outlet temperature is often unmonitored
(as in Estarreja Dow’s cooling tower) the model was trained by
only minimizing the error between model prediction and experi-
mental outlet water temperature. Despite this fact, Fig. 10 shows
that the model output for air outlet temperatures is in good agree-
ment with the experimental values.

3.5. Step 5 – model performance evaluation

To quantitatively evaluate the performance of the model the


following statistical parameters were calculated: correlation coef-
Fig. 5. Objective function evolution for tower R-2 for training [j] and validation
ficient (R) and root mean squared error (RMSE). The correlation
[h] sub-sets.
coefficient provides a variability measure of the data reproduced
when the algorithm returns stage efficiency values higher than 1. in the model and the root mean squared error provides the mea-
For the three case studies presented in this work, applying two sure of residual errors and gives a global idea of the difference be-
equilibrium stages was sufficient to provide a good fit to experi- tween the observed and modeled values. These parameters are
mental data, with both stage efficiencies lower than one. calculated using Eqs. (7) and (8) [22], where Yl is the model output
The algorithm stopped when k = 5, k = 4 and k = 8 for tower R-1, Y, is the average value of the model output, Yb l is the experimental
R-2 and Dow’s, respectively. For tower R-1 and R-2 the algorithm’s value.
exit was activated when the objective function of the validation vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
uPm P
sub-set reached a local minimum while for Dow’s cooling u ðY l  Y 2 Þ  m l¼1 ðY l  Y l Þ
2
R ¼ t l¼1 Pm ð7Þ
tower the algorithm stopped when the difference between two b 2
l¼1 ðY l  Y Þ
consecutive iterations of the training sub-set was lower than the rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
established margin. 1 Xm b Þ2
RMSE ¼ ðY l  Y ð8Þ
m l¼1

Fig. 7. Water outlet temperature model predictions vs. observed value for tower R-1: (a) training [}] + validation [j] and (b) test [N]. Assuming a constant Patm of 101.3 kPa.
480 J.A. Queiroz et al. / Energy Conversion and Management 64 (2012) 473–481

Fig. 8. Water outlet temperature model predictions vs. observed value for tower R-2: (a) training [}] + validation [j] and (b) test [N]. Assuming a constant Patm of 101.3 kPa.

(a) (b)

Fig. 9. Water outlet temperature model predictions vs. observed value for Dow’s tower: (a) training [}] + validation [j] and (b) test [N].

Fig. 10. Air outlet temperature model predictions vs. observed value for training [}], validation [j] and test [N] in: (a) tower R-1 and (b) tower R-2. Assuming a constant Patm
of 101.3 kPa.

Table 5
Model performance parameters for the different data sets. The variable that is being analyzed is the water outlet temperature, Tw out
.

Training Validation Test


R-1 R-2 Dow R-1 R-2 Dow R-1 R-2 Dow
RMSE 0.206 0.172 0.175 0.291 0.129 0.102 0.236 0.218 0.118
R 0.997 0.997 0.984 0.991 0.997 0.996 0.994 0.993 0.996
J.A. Queiroz et al. / Energy Conversion and Management 64 (2012) 473–481 481

In Table 5 model performance parameters regarding outlet with a larger process model enabling a better process control and
water temperature (Tw out) are summarized. For the test data energy management when integrated with the distributed control
sub-sets of tower R-1 and R-2, respectively, RMSEtest is: 0.23; system.
0.218 and Rtest: 0.994; 0.993, as for Dow’s industrial cooling tower
the test sub-set RMSEtest is: 0.129 and Rtest: 0.991. The values of the Acknowledgments
model evaluation parameters indicate a good fit between experi-
mental and model values for both the laboratory setup (tower R- Authors are grateful to Dow Portugal – Produtos Químicos,
1 and R-2) and the industrial application (Dow’s cooling tower). Sociedade Unipessoal, Lda and to Fundação para a Ciência e Tecn-
Although the model was trained using only water outlet tem- ologia, namely for the Ph.D. student support through fellowship
peratures the model performance parameters regarding air outlet SFRH/BDE/51013/2010.
temperatures (Ta out) in tower R-1 and R-2 also indicate an accept-
able fit between model predictions and experimental values: References
RMSEtest = 0.349; Rtest = 0.988 for tower R-1 and RMSEtest = 0.130;
Rtest = 0.997 for tower R-2. In an industrial context air outlet tem- [1] Jin GY, Cai WJ, Lu L, Lee EL, Chiang A. A simplified modeling of mechanical
cooling tower for control and optimization of HAVC systems. Energy Convers
perature is not a critical variable to be controlled, therefore these Manage 2007;48(2):355–65.
values are not available for Dow’s case study and the comparison [2] Benton DJ, Hydeman M, Bowman CF, Miller P. An improved cooling tower
between model and experimental air outlet temperatures was algorithm for the CoolTools™ simulation model. ASHRAE Trans
2002;108(1):AC-02-9-4.
not possible to accomplish. [3] Kloppers JC, Kröger DG. A critical investigation into the heat and mass transfer
All predicted values are within a margin of ±2 % from the exper- analysis of counter flow wet-cooling towers. Int J Heat Mass Transfer
imental values for water outlet temperatures and ±3 % for air outlet 2005;48(3):765–77.
[4] Picardo JR, Variyar JE. The Merkel equation revisited: a novel method to
temperatures (tower R-1 and R-2). compute the packed height of a cooling tower. Energy Convers Manage
2012;57:167–72.
[5] Castro MM, Song TW, Pinto JM. Minimization of operational costs in cooling
water systems. Br J Chem Eng 2000;78(2):192–201.
4. Conclusions [6] Khan JR, Qureshi BA, Zubair SM. A comprehensive design and performance
evaluation study of counter flow wet cooling towers. Int J Refrig
Cooling tower outlet temperature was predicted by applying 2004;27(8):914–23.
[7] Al-Waked R, Behnia M. CFD simulation of wet cooling towers. Appl Therm Eng
the equilibrium stage approach with Murphree stage efficiencies. 2006;26(4):382–95.
Two equilibrium stages were sufficient to provide a good fit for [8] Rubio-Castro E, Serna-González M, Ponce-Ortega JM, Morales-Cabrera MA.
all three case studies. Optimization of mechanical draft counter flow wet-cooling towers using a
rigorous model. Appl Therm Eng 2011;31(16):3615–28.
The proposed approach was validated not only with published
[9] Hosoz M, Ertunc HM, Bulgurcu H. Performance prediction of a cooling tower
data but also with data provided by a real industrial application. using artificial neural network. Energy Convers Manage 2007;48(4):1349–59.
This approach can be useful when simulating a process, allowing [10] Pan TH, Shieh SS, Jang SS, Tseng WH, Wu CW, Ou JJ. Statistical multi-model
approach for performance assessment of cooling tower. Energy Convers
the prompt evaluation of the impacts for different operation
Manage 2011;52(2):1377–85.
conditions. [11] Serna-González M, Ponce-Ortega JM, Jiménez-Gutiérrez A. MINLP optimization
Although only the outlet water temperature was used to iden- of mechanical draft counter flow wet-cooling towers. Chem Eng Res Des
tify model parameters, predicted air outlet temperature was also 2010;88(5–6):614–25.
[12] Rao RV, Patel VK. Optimization of mechanical draft counter flow wet-cooling
in good accordance with experimental values. tower using artificial bee colony algorithm. Energy Convers Manage
A critical element of process simulation is proper physical prop- 2011;52(7):2611–22.
erty methods, and determination of binary interaction parameters. [13] Simpson W, Sherwood TK. Performance of small mechanical draft cooling
towers. ASRE J Refrig Eng 1946;52(6):543–76.
Since ASPEN PLUS is a process simulator that provides several [14] ASPEN PLUS, V7.1. Aspen Technology Inc.; 2008.
built-in model blocks that can be directly applied in process simu- [15] Aktepe A, Öncel Ç, Ersöz S. An artificial neural network model on welding
lation and has an extensive physical property database, it is useful process control of 155 mm artillery ammunition. In: 6th International
advanced technologies symposium (IATS’11), Elazığ, Turkey; 16–18 May 2011.
to apply the methodology described in this work to other process [16] Fang Q. Distinctions between Levenberg–Marquardt method and Tikhonov
units where it is necessary to fit model parameters to experimental regularization. Dartmouth College Publication; 2004.
data for example in distillation towers data reconciliation. For this, [17] Aspen Simulation Workbook, V7.1. Aspen Technology Inc.; 2009.
[18] Aspen Simulation Workbook User Guide, V7.1. Aspen Technology Inc.; 2009.
model parameters such as Murphree or overall column efficiencies
[19] Zubair BA, Qureshi SM. A complete model of wet cooling towers with fouling
are adjusted to a wide variety of operating conditions to produce in fills. Appl Therm Eng 2006;26(16):1982–9.
the best match to plant data. Model parameters that minimize [20] Heidarinejad G, Karami M, Delfani S. Numerical simulation of counter-flow
wet-cooling towers. Int J Refrig 2007;32(5):996–1002.
the overall sum of the squared difference between experimental
[21] Weather Underground; 2011 (May and June). <www.wunderground.com/
and simulated values are then selected. Since model parameters weatherstation/WXDailyHistory.asp?ID=IAVEIROG3>, <www.wunderground.com/
are adjusted to real data, once the proposed methodology is ap- weatherstation/WXDailyHistory.asp?ID=IPORTUGA67> [retrieved 23.02.12].
plied to determine model parameters off-design operating condi- [22] Sousa SIV, Martins FG, Alvim-Ferraz MCM, Pereira MC. Multiple linear
regression and artificial neural networks based on principal components to
tions are accurately simulated using the process simulator predict ozone concentrations. Environ Model Softw 2007;22(1):97–103.
model. Additionally, this approach has the ability to be integrated

You might also like