You are on page 1of 14

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220675196

A stochastic model for performance analysis of pharmaceutical high purity


water systems

Article  in  Simulation Modelling Practice and Theory · February 2012


DOI: 10.1016/j.simpat.2011.10.001 · Source: DBLP

CITATION READS

1 132

3 authors:

Frank Riedewald Edmond Byrne


University College Cork University College Cork
25 PUBLICATIONS   93 CITATIONS    105 PUBLICATIONS   639 CITATIONS   

SEE PROFILE SEE PROFILE

Kevin Cronin
University College Cork
116 PUBLICATIONS   1,110 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Transdisciplinary Perspectives on Transitions to Sustainability View project

Industrial Enzymology View project

All content following this page was uploaded by Edmond Byrne on 30 April 2019.

The user has requested enhancement of the downloaded file.


Simulation Modelling Practice and Theory 21 (2012) 26–38

Contents lists available at SciVerse ScienceDirect

Simulation Modelling Practice and Theory


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

A stochastic model for performance analysis of pharmaceutical high


purity water systems
Frank Riedewald ⇑, Edmond Byrne, Kevin Cronin
Department of Process and Chemical Engineering, Food Science Building, University College Cork, Ireland

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

Article history: The operating performance of high purity water systems such as pharmaceutical Deionized
Received 18 December 2010 and Water for Injection (DI/WFI) distribution systems can be difficult to analyse due to the
Received in revised form 2 October 2011 highly variable demand that is drawn from these systems; a situation compounded by
Accepted 3 October 2011
schedule and demand volume uncertainties. This work presents a stochastic model to sim-
ulate volume and schedule uncertainties of industrial size DI/WFI systems. The model uti-
lises discrete-event simulation combined with the Monte Carlo method to calculate the
Keywords:
demand profile of the distribution system and a continuous simulation to compute the var-
Discrete-event simulation
Stochastic modelling
iation of the level in the storage tank. It is shown that the model may be used to predict if
Water for Injection an existing DI/WFI systems is capable of delivering water for a new process, for which his-
WFI torical data is lacking. The model is programmed in Excel 2003 and is available for down-
Capacity assessment load [1] as open software.
Ó 2011 Elsevier B.V. All rights reserved.

1. Introduction

Parenterals i.e. injectable drugs form a large and important group within the available medical arsenal. Many injectables
e.g. vaccines against tetanus, rubella, measles, flu, mumps or hepatitis require highly purified water, so called Water for
Injection (WFI), as a delivery vehicle for the active ingredient. For a number of reasons purified pharmaceutical grade water
such as Deionised Water (DI) and WFI in particular are the most expensive utility employed in a pharmaceutical plant. First
the capital expenditure is high as high grade stainless steels have to be employed [2,3] to convey the water. Second because
the WFI forms part of the medication it is highly regulated [2,3], requiring companies to satisfy regulatory authorities for
instance the United States FDA (Food and Drug Administration) [2,3] that the DI/WFI system is capable of producing the re-
quired quality water consistently. As a result there is an ongoing costs of quality control, which includes microbiological con-
trol of the water [3] and cost of regulatory scrutiny. Third the production of DI/WFI is expensive as a complicated process
involving a number of steps [6] must be employed. Multi-media filters, ion exchangers, cartridge filters, reverse osmosis
units, continuous electro-deionisation units and distillation are generally employed to achieve the required purity require-
ments of WFI. Usually distillation, an energy intensive and hence costly operation, is the final purification step involved in
the production of WFI. This is because distillation must be used should the medication be for sale in the European market as
stipulated by the European Pharmacopeia (EP) [2,3] (see Table 1). An indication of the purity levels required is given by Ta-
ble 2, which compares the purity of naturally occurring waters to purified water. Furthermore, it is also energy demanding to
distribute WFI within a production facility as the temperature of the WFI in the distribution system is typically maintained at

⇑ Corresponding author. Tel.: +353 086 2463 832; fax: +353 (0)21 4270249.
E-mail address: frankriedewald@gmail.com (F. Riedewald).

1569-190X/$ - see front matter Ó 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.simpat.2011.10.001
F. Riedewald et al. / Simulation Modelling Practice and Theory 21 (2012) 26–38 27

Nomenclature

acti,k opening/closing event of valve i, k (N/A)


i integer parameter (N/A)
k integer parameter (N/A)
Pi variation in percent of full volume (%)
tapi valve (tap) i along the distribution system (N/A)
t close
i;k scheduled closing time for each acti,k (h:m:s)
t close
i;k;new new closing time for each acti,k (h:m:s)
t open
i;k
scheduled opening time for each acti,k (h:m:s)
t open
i;k;new
new closing time for each acti,k (h:m:s)
sim
t simulated time (h:m:s)
tE time at end of simulated time (h:m:s)
t0 time at start of simulated time (h:m:s)
Dtofftake
i;k opening/closing interval for each acti,k (h:m:s)
Dtuncertain
i;k uncertain opening time (h:m:s)
Ui,k binary variable either (1) or (1) (N/A)

V DI;WFI volumetric DI flowrate to WFI generation plant (m3/h)
 
V WFI;1 to V WFI;n offtake from WFI loop (m3/h)

V WFI;Blow WFI generation blowdown (%)

V WFI;Gen volumetric flowrate WFI from WFI generation plant (m3/h)

V WFI;Gen;max maximum volumetric flowrate from WFI generating plant (m3/h)
VWFI,L water volume in the WFI storage tank (m3)
VWFI,L,C temperature compensated water volume in WFI storage tank (m3)
VWFI,L,max maximum allowable water volume in WFI storage tank (m3)
VWFI,L,min minimum allowable water volume in the WFI storage tank (m3)
VWFI,L,St start volume in WFI storage tank at t0 (m3)

V WFI;off volumetric off take from the WFI system (m3/h)

V WFI;off i;k volumetric offtake for each acti;k (m3/h)

3
V actual
WFI;off;i;k volumetric flowrate for each acti,k (m /h)

3
DV uncertain
WFI;off;i;k stochastic volumetric flowrate variation for each acti,k (m /h)

V WFI;Pump WFI distribution pump delivery volume (m3/h)
X random variable, uniformly distributed in {0, . . ., 1} (N/A)
Yi generic stochastic variable (N/A)
Zi,k generic stochastic variable (N/A)

Greek variables
q20 water density at 20 °C: q20 = 998.21 kg/m3 (kg/m3)
q80 water density at 80 °C: q80 = 971.79 kg/m3 (kg/m3)

temperatures exceeding 70 °C and controlled within a temperature band of ±5.0 °C in order to suppress bacterial growth
[2,4].
Many different DI/WFI distribution configurations are used in industry. The literature describes distribution systems with-
out any buffer tank [7], with numerous tanks along the distribution system [3], batch systems [3] among others. The vast
majority of all DI/WFI systems in industry is, however, the looped system with a storage tank located in each of the DI
and WFI distribution loops as schematically shown in Fig. 1 [2,3]. The reason for the success of this DI/WFI configuration
is that the hydraulic conditions and temperature profiles along the entire distribution system can readily be achieved.
There is no general agreement on some design details of pharmaceutical DI/WFI systems, such as the requirements for
minimum water velocities [2,3] or the value of reducing internal pipe roughness [3]. Typical design parameters of DI/WFI
systems as applied in practice are summarised in Table 3. One important design parameter for DI/WFI systems is reliability,
as physical breakdowns of DI/WFI systems may result in prolonged and costly production shutdowns. Nevertheless nowa-
days DI/WFI systems can be built which may operate continuously for many years [7]. This work is not concerned with any of
these design parameters however, as these are discussed in detail by, for instance, Meltzer [2] or the ISPE Baseline Guide on
28 F. Riedewald et al. / Simulation Modelling Practice and Theory 21 (2012) 26–38

Table 1
Comparison of the stipulations of the US and European Pharmacopeia on WFI [2,6].

US Pharmacopeia (USP) European Pharmacopeia (EP)


Source water Drinking water Drinking water or purified water
Preparation Distillation or other suitable process Distillation
Total organic 0.5 ppm 0.5 ppm
carbon (TOC)
Conductivity 1.1 lS 20 °C 1.1 lS 20 °C
Appearance N/A Clear and odourless
Bacterial Not more than 0.25 USP EU/mL (EU: endotoxin unit, measure of toxin Less than 0.25 IU/mL (IU: international unit, measure
endotoxins released upon cell destruction) of biological activity)
Bacterial count 10 CFU/100 mL (CFU: colony forming units, measure of cell activity) 10 CFU/100 mL

Table 2
Water impurities and water purity for various waters (adapted from Dabbah [5]).

Types of water Impurities (ppm) Purity (%)


Sea water 30,000 97
Potable water 500 99.95
Mountain water 50 99.995
Purified water 1 99.9999

Fig. 1. Process flowsheet of a typical DI/WFI distribution system.

Water Systems [3]. Instead it is concerned with the effective sizing of a DI/WFI system i.e. the size of the DI/WFI generation
and distribution system including the storage tank volume as a function of expected demand.
The suitability of an existing DI/WFI system to accommodate a different, maybe new process can be challenging to answer
with confidence, because it is difficult to analytically predict the highly variable demand profile a utility system may have to
provide. This difficulty is compounded by uncertainties such as schedule and volume uncertainties.

Table 3
Typical design parameters of industrial size DI/WFI systems [3].

Parameter DI system WFI system


Min. fluid velocity in the distribution pipe Not specified to min. 1 m/s min. 1 m/s
Material of construction 316 LSS (LSS: low carbon stainless steel) 316 LSS
Surface finish 0.8–0.5 lm Ra (Ra: average roughness) 0.5–0.25 lm Ra or better and electropolished
Operating temperature Ambient 70–80 °C
Operating pressure (pipe) Above ambient at all times Above ambient
F. Riedewald et al. / Simulation Modelling Practice and Theory 21 (2012) 26–38 29

In practice a diversity factor, denoted fDiv, defined as the quotient of design flow over maximum possible flow, is com-
monly used to handle uncertainty in DI/WFI systems [3]. For unknown reasons this factor is usually between 0.7 and 0.8
[8]. McQueen et al. [9] make the case how woefully inadequate the diversity factor method is to obtain the design load of
an electrical distribution system, as it cannot model the dynamics of the system. Extending their argumentation to DI/
WFI systems makes it clear that a diversity factor should not be used to calculate the design load of a DI/WFI system.
This paper utilises discrete-event simulation to compute how the DI/WFI demand evolves over time. Furthermore this
paper utilises stochastic methods to simulate demand volume and schedule uncertainties in DI/WFI systems. This approach
overcomes the shortcomings of the diversity factor method and allows the calculation of the DI/WFI demand profile while
also allowing experimentation with the DI/WFI system in a virtual world.
Thus far only deterministic models of DI/WFI systems have been described in the literature. Saraph [10] describes a deter-
ministic simulation model of a DI/WFI system. Computer experiments allowed Saraph [10] to show that a recurrent WFI
shortage was a result of poor scheduling of water consumption from different WFI consumers, rather than a water-gener-
ating problem. Alexander [11] assessed how an existing WFI system could cope with projected future demand. Unfortunately
deterministic models may not describe reality very well, as uncertainty is an integral part of any production environment
[12]. Including uncertainty as proposed in this paper should therefore lead to an improved model of a DI/WFI system.

2. Mathematical formulation of the DI/WFI model

A general simplification can be made in the presentation of the model. The mathematical model of the DI and WFI gen-
eration and distribution systems are geometrically similar, as is evident from the process flowsheet (Fig. 1). For this reason,
only the model of the WFI system is developed here.
Three limitations of the model are noted. First, the hydraulics of the DI/WFI system is not modelled. Therefore, the phys-
ical configuration of the DI/WFI distribution system, be it the typical loop, a deadleg or another system [3] is of no concern
here, as the proposed model ignores the layout of the distribution pipe, because it is irrelevant for the computation of the
demand profile. A second limitation of the model is that breakdowns of the DI/WFI system are excluded from modelling,
for DI/WFI systems are generally highly reliable [7]. A third limitation is that interdependencies between the various WFI
demands are not modelled.
The WFI loop contains n offtake valves denoted tapi with i = 1, 2, . . ., n (see Fig. 2 for nomenclature). Each tapi is associated
with k = 1, 2, . . ., m activities denoted acti,k comprised of two events: opening and closing of the valve in question. The open-
ing/closing interval denoted Dt offtake
i;k associated with each acti,k is composed of an opening time denoted topen
i;k and a closing
close
time denoted t i;k as depicted graphically in Fig. 3.
The mathematical model of the tap opening is subject to a number of constraints:

t i,open
k
t i,close
k
t i,open
k +1 t i,close
k +1

Water Tap tap i Water Tap tap i

open open
act i, k act i, k +1
offtake offtake
t i, k t i, k +1

t iopen t iclose t iopen


+1, k +1 t iclose
+1, k +1
+1, k +1, k

Water Tap tap Water Tap tap i+1


i+1

open open
act i +1, k act i +1, k +1
offtake
offtake
t i+ t i +1, k +1
1, k

Simulated
Time tsim

Fig. 2. Nomenclature for the DI/WFI discrete-event simulation.


30 F. Riedewald et al. / Simulation Modelling Practice and Theory 21 (2012) 26–38

t i,open
k t i,open
k, new t i,close
k t i,close
k, new
t i,open open
k +1 t i, k +1, new
t i,close close
k +1 t i, k +1, new

Water Water Tap Open Water Water


Tap Open
Tap Open

Time
t i,offtake
k
t i,offtake
k +1

t i,offtake
k
t i,offtake
k +1

t i,uncertain
k ⋅ Z i, k ⋅ U i, k t i,uncertain
k +1 ⋅ Z i, k +1 ⋅ U i, k +1
Fig. 3. Nomenclature for the stochastic schedule uncertainty variation. The solid lines indicate the crisp schedule, while the dotted lines indicate the
variation of the crisp schedule by the stochastic distribution.

1. Allocation constraints: A tap can only serve one process at any one time; it is not possible to open a tap while it is still
serving another task.
2. Precedence constraints: This constraint ensures the correct sequence of operations of acti,k, acti,k+1, . . ., acti,k+m. The tap close
time tclose
i;k of acti;k must precede the opening time t open
i;kþ1
of acti;kþ1 defined by:
tclose
i;k  t open
i;kþ1 8 i ¼ 1; 2; . . . ; n; 8 k ¼ 1; 2; . . . ; m ð1Þ

Furthermore the open time must proceed the close time for each acti,k:

topen
i;k  t close
i;k 8 i ¼ 1; 2; . . . ; n; 8 k ¼ 1; 2; . . . ; m ð2Þ

3. Volume constraints: The volumetric

offtake

V WFI;off i;k from the distribution loop for each acti,k is assumed to be constant for
a given i and all k reducing V WFI;off;i;k to V WFI;off;i : This simplification is justified by the fact, that a WFI distribution system is
a pumped system operating under a relatively constant pressure.
4. Duration constraints: For each acti,k the time interval Dtofftake i;k ¼ tclose
i;k  topen
i;k is assumed to be constant, not being subject to
uncertainty. This is a reasonable assumption, as only validated volumes VWFI,off,i,k are dispensed from the WFI distribution
system.
5. Non-negativity constraints: None of the input data or the calculated values must assume negative values.

The model is using a step width of one second moving the simulated

time denoted tsim forward. Once the simulated time
has moved forward by one second, the water offtake denoted V WFI; off from the WFI distribution system for every time frame
tsim can finally be calculated as:
 X
n 
V WFI;off ¼ V WFI;off;i 8 tsim ¼ 1; 2; . . . ; 86; 400 ð3Þ
i¼1

with tsim = 86,400 representing a 24-h period expressed in seconds.


Two different uncertainties are chosen to be modelled; (a) dispensed volume and (b) schedule uncertainty:

a. Dispensed volume uncertainty is the uncertainty associated with the amount of water dispensed from the distribution
system for each acti,k.
b. Schedule uncertainty is the uncertainty associated with the opening and closing times of each acti,k, as it would be in a
real production environment [12].

2.1. Stochastic model of tap opening/closing times

Let Zi,k (see Fig. 3 for nomenclature) denote a variable expressing the stochastic distribution of the uncertain opening time
denoted Dtuncertain
i;k for each acti,k. Furthermore let Ui,k be a variable representing the chances of Dt uncertain
i;k be negative or non-
negative, as it is possible for the new opening times denoted topen i;k;new of each acti,k to be earlier or later than originally sched-
uled. Hence the new opening time t open i;k;new
is calculated as:

topen open
i;k;new ¼ t i;k þ Dt uncertain
i;k  Z i;k  U i;k 8 i ¼ 1; 2; . . . ; n 8 k ¼ 1; 2; . . . ; m ð4Þ
F. Riedewald et al. / Simulation Modelling Practice and Theory 21 (2012) 26–38 31

A number of simplifications are made to reduce the dimensions of Zi,k and Ui,k:

1. Each Zi,k could be described by many statistical distributions [13]. The model as can be downloaded [1] uses three statis-
tical distributions namely the Normal, the Beta and the uniform distribution [1]. In addition, the parameters of the three
distributions are, if applicable, constant throughout the simulation, reducing the computational burden.

It is important to allow the user to choose between different stochastic distributions, as measurement uncertainty, for in-
stance, may not always be best described by the Normal distribution [14]. It is, however, a widely used distribution in sta-
tistics [15] and is therefore included in the model. The uniform distribution is often used if no information on the shape of
the distribution is available [15]. The beta distribution is included in the model as it can easily adopt different shapes. Of
course other distributions can be added to the model if necessary.

2. Each Ui,k is reduced to a binary variable, which can be either negative or non-negative depending on the result of the uni-
form probability distribution denoted U(X) as follows:

1; if UðXÞ 6 0:5
U i;k ¼ ð5Þ
1; if UðXÞ > 0:5
with X denoting a random variable.

Further, it is assumed, that the tap opening intervals Dtofftake


i;k are constant and that only the opening topen
i;k and, with it, the
closing times t close
i;k for each acti,k will shift forwards and backwards in time. Hence, the new closing time tclose
i;k;new for each acti,k
can be determined by:
open
tclose offtake
i;k;new ¼ t i;k;new þ Dt i;k 8 i ¼ 1; 2; . . . ; n 8 k ¼ 1; 2; . . . ; m ð6Þ

The precedence constraints presented above must also be observed by the stochastic model if the schedule is subject to sto-
chastic uncertainties. These uncertainties in time will shift the opening topen i;k and closing times t close
i;k backwards or forwards in
open close
time, creating new opening t i;k;new and closing times ti;k;new , which must not clash with events in an earlier or later timeframe as
schematically shown in Fig. 4 or mathematically:

topen close
i;k;new  t i;k;new 8 i ¼ 1; 2; . . . ; n; 8 k ¼ 1; 2; . . . ; m ð7Þ

To avoid a possible violation of the precedence constrain, the VBA program [1] of the DI/WFI system checks the event table
for any possible violations of the precedence constraints taken the maximum possible time variation into account before
proceeding with the simulation. Should a possible violation be found, the user is notified of the location of the possible vio-
lation on the event table and the program terminates.

2.2. Stochastic model of dispensed volume



Let Yi denote a variable describing the stochastic distribution of the dispensed uncertain WFI volume denoted DV uncertain
WFI;off;i;k
for each acti,k. Let Pi denote the maximum variation in percent from full volume. Furthermore let Ui,k be a variable represent-

t open
i +1, k t iopen close
+1, k +1 t i +1, k t close
i +1, k +1

act i +1, k
act i +1, k +1

t offtake
i +1, k +1
t offtake
i+1, k

Simulated
Time
tsim
Fig. 4. Violation of the precedence constraint: the patterned area shows a non-allowable overlap of water tapiþ1 serving two different events actiþ1;k and
actiþ1;kþ1 at the same time.
32 F. Riedewald et al. / Simulation Modelling Practice and Theory 21 (2012) 26–38

 
ing the chances of DV uncertain
WFI;off;i;k being negative or non-negative,

as it is possible that the flowrate V WFI;off;i of each acti,k is more or
less than nominal. Then, the flowrate uncertainty DV uncertain
WFI;off;i;k for each acti;k is calculated as:
 
DV uncertain
WFI;off;i;k ¼ V WFI;off;i  Y i  U i;k  P i =100 8 i ¼ 1; 2; . . . ; n 8 k ¼ 1; 2; . . . ; m ð8Þ

Note that V WFI;off;i , !i and

P i are independent of k due to the volume constraint.
The actual flowrate V actual WFI;off;i;k , that is the flowrate including flowrate uncertainty, for each acti;k is computed as:
  
V actual uncertain
WFI;off;i;k ¼ V WFI;off;i þ DV WFI;off;i;k 8 i ¼ 1; 2; . . . ; n 8 k ¼ 1; 2; . . . ; m ð9Þ

Similar to the tap opening/closing times described above, a number of simplifications are made to reduce the dimensions of
!i and U i;k :
1. Each Z i;k could be described by many statistical distributions [13]. The same comments made for Z i;k apply to Yi.
2. Each U i;k is reduced to a binary variable (see Eq. (5)).

2.3. Model of the WFI storage tank and generating system



Once the water demand V WFI;off from the WFI distribution system for all timeframes tsim has been computed, the variation
of the water level in the storage tank denoted V WFI;L over a 24-h period can be computed from a mass-balance (see Fig. 5) of
the WFI storage tank as follows:
dVWFI;L  
¼ V WFI;L þ ðV WFI;Gen  V WFI;off Þ  dt ð10Þ
dt
with the initial condition of the WFI storage tank volume at t = t0 as:
V WFI;L ¼ V WFI;L;St ð11Þ
The following assumptions are made to arrive at Eq. (10):

1. Water is incompressible; therefore the water removal from the distribution system equals the water removal from the
storage tank.
2. The water temperature of the WFI system is 80 °C and controlled within a temperature band of ±5.0 °C in order to main-
tain a low bacterial count [2,4]. Consequently, the water density along the WFI distribution system is assumed to be
constant.

The differential equation (10) is integrated by the Euler method, which according to Luyben and Luyben [16] is a suitable
method for this problem. The integration step is, in line with the step width of the discrete event simulation, one second.
Furthermore the water level in the WFI storage tank is bounded between a minimum and maximum volume:
V WFI;L;min 6 V WFI;L 6 V WFI;L;max ð12Þ
The raw water for the WFI generation plant is provided by the DI loop. A mass-balance on the WFI generation plant (see
Fig. 6) provides the amount of DI water required as:
  
V DI;WFI ¼ V WFI;Gen þ V WFI;Blow ð13Þ

V WFI;Blow in the above equation denotes the amount of water from the WFI generation plant going to blowdown; here as-
sumed to go to drain. The blowdown of a WFI plant is usually between 5% and 8% of the raw (DI) water [2] and is another
input variable of the model.

The flow of water V WFI;Gen from the WFI generation plant to the WFI storage tank is controlled to maintain the level in the
WFI storage tank at its maximum allowable level V WFI;L;max . The control is assumed to be perfect; any delay, overshooting or

V WFI, Gen •
V WFI, 1

V WFI, 2
WFI Water
Storage Tank


VWFI, L V WFI, n

Fig. 5. Mass balance of the WFI-storage tank.


F. Riedewald et al. / Simulation Modelling Practice and Theory 21 (2012) 26–38 33

• •
V DI, WFI WFI Generation Plant V WFI, Gen


V WFI, Blow
Fig. 6. Mass balance of the WFI-generation plant.

dampening effects are ignored. The water flow from the WFI generation plant replenishes the level in the WFI storage tank in
any timeframe tsim depending on the offtake from the WFI loop according to four rules:
 
1. Should the WFI offtake from the distribution loop V WFI;off be less than the capacity of the WFI generation plant V WFI;Gen and
the water level V WFI;L in the WFI storage tank is as its maximum, then the level in storage tank will be maintained by the
water flow from the WFI generation station:
 
V WFI;Gen ¼ V WFI;off if V WFI;L ¼ V WFI;L;max 8 t sim ¼ 1; 2; . . . ; 86; 400 ð14Þ
2. Should the water level V WFI;L in the WFI storage tank drop below the maximum allowable value, then the flowrate from
the WFI generation plant shall increase to its maximum value to replenish the water in the storage tank to its maximum
level as soon as possible:
 
V WFI;Gen ¼ V WFI;Gen;max if V WFI;L < V WFI;L;max 8 tsim ¼ 1; 2; . . . ; 86; 400 ð15Þ
3. Should the water level VWFI,L in the WFI storage tank exceed the maximum allowable value, then the flowrate from the
WFI generation system shall be zero and the level in the storage tank shall be the maximum allowable level:

V WFI; Gen ¼ 0 if V WFI;L ¼ V WFI;L;max 8 t sim ¼ 1; 2; . . . ; 86; 400 ð16Þ
4. Should the water level V WFI;L in the WFI storage tank drop below the minimum allowable level V WFI;L;min , the simulation
terminates as a non-allowable situation arose.

It is assumed that the production process requires WFI at 20 °C and that the volumetric measurement of the amount of
WFI dispensed is also made at 20 °C. The WFI is, however, stored and distributed at 80 °C to maintain a low bacterial count
[2], resulting in a temperature difference between the storage and distribution system and the raw and dispensed water. This
temperature difference requires a density compensation of the amount of water in the WFI water storage tank in order to
obtain a temperature compensated WFI water volume as follows:
q20
V WFI;L;C ¼ V WFI;L ð17Þ
q80

3. Methods

3.1. Methods used to solve the model

The method applied to solve the model comprises of two steps: discrete-event and continuous simulation. Discrete Event
Simulation (DES) is used to solve the problem posed in this work. DES is often used to gain insights into complex systems,

Start Event Stop Event


Valve i Opened Valve i Closed

Water Tap Water Tap Water Tap


Open Open Open

Time

Activity Wait time Activity


between
Activities

Fig. 7. On the workings of a discrete-event simulation.


34 F. Riedewald et al. / Simulation Modelling Practice and Theory 21 (2012) 26–38

which elude analytical treatment. It is most applicable for systems in which the state of the systems changes at discrete
times, such as car arrivals at traffic intersections [17] or indeed WFI systems [10,11]. Specifically, in WFI systems the state
of the system changes should a valve on the distribution loop open or close as is graphically shown in Fig. 7.
The stochastic model is solved with the Monte Carlo (MC) method [15,18]. Once the MC method has solved the water
demand profile a continuous simulation can be used to calculate the fluctuations of the level in the WFI tank and the re-
quired WFI generation capacity.

3.2. Model validation

Model validation is a critical part of the model building process, proving the suitability of a model for the problem at
hand. Validating a model against a real system provides most confidence into a model, but it is not always possible, as
the system may not exist as yet or data may be too difficult to obtain [17–19]. Unfortunately a real DI/WFI system is not
available for this work, thus requiring different techniques to validate the model.
Sargent [20] describes fifteen different techniques to model validation, of which four were used to validate the model.
Firstly the Internal Validity test was applied to the stochastic model to ensure consistency of the results. Should the results
lack consistency the results may be questionable. Secondly the relationship between input and outputs was checked if rea-
sonable (Face Validity test). Thirdly the graphical outputs were checked if reasonable (Operational Graphics test). The last test
used is the Traces test, which follows specific inputs trough the simulation to check if the logic of the model is correct. The
model passed these four validation tests, and is therefore considered to be accurate for the purposes of this work.

3.3. Excel as a simulation platform

Excel was chosen over commercially available discrete-event simulation packages to provide a cost effective, portable,
and open model available for download [1]. The model is programmed in Visual Basic for Applications (VBA), which is a ob-
ject-orientated general purpose programming language included in the standard version of Excel 2003 [21]. The spreadsheet
grid serves as the input/output interface allowing users unfamiliar with VBA to execute DI/WFI simulations.

4. Illustrative example

The following example demonstrates how the model can be applied on a WFI system as depicted on the right hand side of
Fig. 1. The objective is to determine if a future production process, dubbed ‘‘New Process’’ can be supported by an existing
WFI system. Let’s assume that the company’s existing process, named ‘‘Existing Process’’, was run for a number years and
that it was used to validate the model. Furthermore the model verified operational experience of intermittent WFI shortages
due to the WFI demand exceeding the WFI pump delivery capacity. This effect can be seen in the right hand panels of Fig. 8,
3rd and 4th MC Run, displaying WFI demand peaks exceeding the WFI pump flowrate indicated by the dotted line. During
times of peak demand, WFI process operations fed from WFI valves located towards the end of the distribution loop are
starved of WFI resulting in delays of these process operations till the overall demand has again dropped below 43 m3/h.
The left hand panels of Fig. 8 display the water level in the storage tank over the simulated day with the dotted line indi-
cating the minimum allowable volume in the storage tank, which is predicted not to be reached. Let’s further assume that
the production schedule of the ‘‘New Process’’ is relatively uncertain as it has never been produced on a large scale, making it
important to allow for volume and schedule uncertainties. Because the process is new and uncertain the uniform distribu-
tion as the distribution of least knowledge [15,18] is used to model all uncertainties. In addition the WFI demand from the
‘‘New Process’’ is such that the majority of the daily WFI demand occurs in the afternoon increasing the likelihood that the
WFI storage tank might be empty resulting in lengthy production delays. Therefore the company must demonstrate that the
existing WFI system can support the ‘‘New Process’’ in order to be considered to produce the new drug.
The process conditions of the WFI system are as follows:

1. The WFI storage tank volume is 40 m3.


2. The WFI generating capacity is 20.6 m3/h.
3. About 50% of the tasks are manual filling operations i.e. parts washing and line rinsing, which are having a volumetric
accuracy of ±10% of full volume. All other dispensing operations are automated having a volumetric accuracy of ±1% of
full volume.
4. The WFI generation blowdown to drain is 5%.
5. The minimum allowable water volume in storage tank is 2 m3.
6. The WFI distribution pump PU-101 has a delivery capacity of V WFI;Pump ¼ 43 m3 =h.

Should the simulation show that the ‘‘New process’’ cannot be accommodated by the WFI system the WFI storage capac-
ity, the WFI generation capacity or both can be increased. The delivery capacity of the WFI distribution pump can, however,
not be increased as otherwise the pump pressure head would get too high.
F. Riedewald et al. / Simulation Modelling Practice and Theory 21 (2012) 26–38 35

Fig. 8. Graphical results ‘‘Existing Process’’, 1st–6th MC Run.

Fig. 9 displays how the water level in the storage tank and the water demand profile for the ‘‘New Process’’ is predicted to
behave for six different MC runs out of a total of 50 MC runs. Whereas the water level in the storage tank is predicted to be
lower than in the ‘‘Existing Process’’, it is not predicted to reach the minimum allowable level. Therefore the tank and water
generation systems are suitable to support the new process. But similar to the existing process there are production delays
predicted caused by peak

demands in WFI (see Fig. 9, 2nd MC run). The chances of how often the WFI demand may exceed
the delivery capacity V WFI;Pump of the WFI distribution pump may, however, be difficult to assess from a graphical output as
36 F. Riedewald et al. / Simulation Modelling Practice and Theory 21 (2012) 26–38

Fig. 9. Graphical results ‘‘New Process’’, 1st–6th MC Run.

given in Fig. 9. But presenting these times as a probability density function (PDF) generated from all 50 MC runs displaying
the numerical measure ‘‘Minutes per day the WFI demand exceeds the WFI distribution pump PU-101 flowrate of 43 m3/h’’
may give an improved presentation.
Comparing the PDFs of the ‘‘Existing Process’’ (see Fig. 10a) with the ‘‘New Process’’ (see Fig. 10b) shows that the existing
WFI system is capable of supporting the ‘‘New Process’’ because the predicted production delays of the ‘‘New Process’’ are
slightly fewer than the delays occurring by the ‘‘Existing Process’’. For both processes the majority of WFI starvation time
intervals are of short duration (<5 min) and the WFI starvation times exceeding 15 min are of low probability.
F. Riedewald et al. / Simulation Modelling Practice and Theory 21 (2012) 26–38 37

Fig. 10. PDFs of WFI starvation of the stochastic simulations of the: (a) ‘‘Existing Process’’ and (b) ‘‘New Process’’.

Fig. 11. PDFs of WFI starvation of the stochastic simulations of the: (a) ‘‘New Process’’ 1% volume uncertainty and (b) ‘‘New Process’’ 10% volume
uncertainty.

A sensitivity analysis can also be performed on the proposed model. Fig. 10 shows a comparison between two experi-
ments. In the first experiment all volume uncertainty inputs are set to ±1% (Fig. 11a) and in the second all volume uncer-
tainty inputs are set to ±10% (Fig. 11b). Because the PDFs are virtually the same, the influence of volume uncertainty may
be ignored; reducing the data accuracy requirements for a simulation of this system.

5. Conclusion

This work presents an industrial scale stochastic model to simulate volume and schedule uncertainties of pharmaceutical
DI/WFI systems. As shown the model can, for instance, be used to assess if an existing WFI systems is capable of delivering
water to a new process for which historical may be lacking. Likewise the model may be used in the design of new DI/WFI
systems. Furthermore the performance of Ultrapure Water systems as employed in the microelectronics industry may be
assessed with the model. Other chemical plant utilities, such as clean stream, cooling water, heat transfer media systems
may also be assessed with the principles outlined in this paper. In such applications, however, the model may have to be
modified. The model may also be extended to include more complex DI/WFI distribution systems for example such in which
two or more storage tanks are placed in different locations along a DI/WFI distribution loop.

References

[1] Download DI/WFI Stochastic Excel Simulation Spreadsheet, Department of Process & Chemical Engineering, University College Cork, 2011, <http://
www.ucc.ie/processeng/links/utility/stochastic>.
[2] T.H. Meltzer, High-Purity Water Preparation: For the Semiconductor, Pharmaceutical, and Power Industries, second ed., Tall Oaks Publishing, Littleton,
Colorado, 1997.
[3] Anonymous, Water and Steam Systems, first ed., International Society for Pharmaceutical Engineering (ISPE), 2001.
38 F. Riedewald et al. / Simulation Modelling Practice and Theory 21 (2012) 26–38

[4] J.E. Martinez, Hyperthermophilic microorganisms and USP hot water systems, Pharmaceutical Technology (2004) 50–65.
[5] R. Dabbah, USP pharmaceutical waters, Part 1 bulk waters, BioProcess International (2006) 50–55.
[6] R. Dabbah, USP pharmaceutical waters, Part 2 packaged waters and harmonization issues, BioProcess International (2006) 58–62.
[7] B.H. Junker, M. Stanik, J. Adamca, K. LaRiviere, M. Abbatiello, P. Salmon, An ambient water loop system for USP purified water, Bioprocess Engineering
17 (5) (1997) 277–286.
[8] M. Wu, D. Sun, J.H. Tay, Development of a practical model for capacity evaluation of ultrapure water systems, Desalination 161 (3) (2004) 223–233.
[9] D.H.O. McQueen, P.R. Hyland, S.J. Watson, Monte Carlo simulation of residential electricity demand for forecasting maximum demand on distribution
networks, IEEE Transaction on Power Systems 19 (3) (2004) 1685–1689.
[10] P.V. Saraph, Biotech industry: simulation and beyond, in: B.A. Peters, J.S. Smith, D.J. Medeiros, M.W. Rohrer (Eds.), Proceedings of the 2001 Winter
Simulation Conference, 2001, pp. 838–843.
[11] C.W. Alexander, Discrete event simulation for batch processing, in: L.F. Perrone, F.P. Wieland, J. Liu, B.G. Lawson, D.M. Nicol, R.M. Fujimoto (Eds.),
Proceedings of the 2006 Winter Simulation Conference, 2006, pp. 1929–1934.
[12] S.J. Honkomp, S. Lombardo, O. Rosen, J.F. Pekny, The curse of reality – why process scheduling optimization problems are difficult in practice,
Computers & Chemical Engineering 24 (2-7) (2000) 323–328.
[13] M. Evans, N. Hastings, B. Peacock, Statistical Distributions, second ed., John Wiley & Sons, New York, 1993.
[14] M. Drosg, Dealing with Uncertainties: A Guide to Error Analysis, second ed., Springer, Dordrecht, Heidelberg, London, New York, 2009.
[15] K. Cronin, J.P. Gleeson, M.S.R. Shyam, S. Sablani, Ashim K. Datta, Arun S. Mujumdar (Eds.), ‘‘Monte Carlo Simulation,’’ Handbook of Food and Bioprocess
Modeling Techniques, Boca Raton, CRC Press, 2007.
[16] M.L. Luyben, W.L. Luyben, Essentials of Process Control, McGraw-Hill, New York, 1993.
[17] Handbook of Simulation: Principles, Methodology, Advances, Applications, and Practice, John Wiley & Sons, Inc., New York, 1998.
[18] A.M. Law, D.M. Kelton, Simulation Modeling and Analysis, second ed., McGraw-Hill, New York, 1991.
[19] A.M. Law, How to build valid and credible simulation models, in: M.E. Kuhl, N.M. Steiger, F.B. Armstrong, J.A. Joines (Eds.), Proceedings of the 2006
Winter Simulation Conference, 2006, pp. 24–32.
[20] R.G. Sargent, Verification and validation of simulation models, in: S.G. Henderson, B. Biller, M.-H. Hsieh, J. Shortle, J.D. Tew, R.R. Barton (Eds.),
Proceedings of the 2007 Winter Simulation Conference, 2007, pp. 124–137.
[21] J. Green, S. Bullen, R. Bovey, M. Alexander, Excel 2007 VBA Programmers Reference, Indianapolis, Wiley Publishing, Inc., 2007.

View publication stats

You might also like