You are on page 1of 16

Chemical Engineering Communications

ISSN: 0098-6445 (Print) 1563-5201 (Online) Journal homepage: https://www.tandfonline.com/loi/gcec20

A Hybrid Method for Stochastic Performance


Modeling and Optimization of Chemical
Engineering Processes

Usman Abubakar , Srinivas Sriramula & Neill C. Renton

To cite this article: Usman Abubakar , Srinivas Sriramula & Neill C. Renton (2015)
A Hybrid Method for Stochastic Performance Modeling and Optimization of Chemical
Engineering Processes, Chemical Engineering Communications, 202:2, 217-231, DOI:
10.1080/00986445.2013.838159

To link to this article: https://doi.org/10.1080/00986445.2013.838159

Published online: 15 Sep 2014.

Submit your article to this journal

Article views: 405

View related articles

View Crossmark data

Citing articles: 1 View citing articles

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=gcec20
Chemical Engineering Communications, 202:217–231, 2015
Copyright # Taylor & Francis Group, LLC
ISSN: 0098-6445 print/1563-5201 online
DOI: 10.1080/00986445.2013.838159

A Hybrid Method for Stochastic Performance Modeling


and Optimization of Chemical Engineering Processes
USMAN ABUBAKAR, SRINIVAS SRIRAMULA, and NEILL C. RENTON
School of Engineering, University of Aberdeen, Aberdeen, United Kingdom

As chemical engineers seek to improve plant safety, reliability, and financial performance, a wide range of uncertaintyladen deci-
sions need to be made. It is widely agreed that probabilistic approaches provide a rational framework to quantify such uncertainties
and can result in improved decision making and performance when compared with deterministic approaches. This article proposes
a novel method for design and performance analysis of chemical engineering processes under uncertainty. The framework combines
process simulation tools, response surface techniques, and numerical integration schemes applied in structural reliability problems
to determine the probability of a process achieving a performance function of interest. The approach can be used to model pro-
cesses in the presence or absence of performance function(s), with or without parameter interactions, at both design and oper-
ational phases. With this, process behavior can be quantified in terms of stochastic performance measures such as reliability
indices and the associated most probable process design=operating conditions, providing a simple way to analyze a wide range
of decisions. To validate the applicability of the proposed framework, three case study systems are considered: a plug flow reactor,
a heat exchanger, and finally a pump system. In each case, performance criteria based on the original physical model and the sur-
rogate model are set up. Reliability analysis is then carried out based on these two models and the results are assessed. The results
show that the proposed framework can be successfully applied in chemical engineering analysis with additional benefits over the
traditional deterministic methods.
Keywords: Optimization; Process modeling; Reliability analysis; Response surface methodology; Stochastic programming;
Uncertainty quantification

Introduction and Wozny, 2009; Ascough II et al., 2008; Badar et al.,


1993; Cheng and Sandu, 2009; Diwekar and Rubin, 1991).
Uncertainties in process plant behavior continue to cause Using deterministic methods, low probability events (which
considerable concern to chemical engineers involved in their could be of high consequence) and random system behavior
design, commissioning, and operation. In the presence of may not be adequately accounted for by just imposing
uncertainties stemming from lack of information and=or design factors on a less informed empirical basis. As
random physical variation, there is still a need to determine a result, plant equipment and systems can be over- or
optimum design factors, production=financial performance, under-designed, with potentially serious financial, safety,
and overall system reliability, and ensure that the plant is and environmental consequences. These issues continue to
operating in a safe manner under a variety of operating emphasize the need for stochastic approaches for systems
conditions (Lasserre and Roubellat, 1985; Pistikopoulos design and performance analysis.
and Grossmann, 1988; Pistikopoulos and Mazzuchi, 1990; To address this, considerable effort has been committed
Szépvölgyi, Jr., et al., 1999; Vasquez and Whiting, 2004). to the development of stochastic analysis techniques in a
Traditional deterministic approaches are limited in their range of engineering disciplines, particularly in the field of
ability to give the engineer insight to the behavior of com- structural engineering (Melchers, 1999; Zio, 2009). The
plex chemical engineering systems featuring high dimension- application of stochastic methods in chemical engineering
ality, nonlinear performance functions, and random analysis has also received attention. A novel approach for
behavior. Unfortunately such conditions are often present flexibility analysis for processes under uncertainty was
in modern process engineering systems (Arellano-Garcia proposed by Pistikopoulos and Mazzuchi (1990). In this
approach, the process was assumed to be governed by a
Address correspondence to Srinivas Sriramula, Lloyd’s Register linear model and the uncertain parameters were assigned
Foundation (LRF) Centre for Safety & Reliability Engineering, Gaussian distribution functions to compute the performance
School of Engineering, University of Aberdeen, Aberdeen, index. Badar et al. (1993) investigated the performance of
AB24 3UE, UK. E-mail: s.sriramula@abdn.ac.uk a heat exchanger under uncertainty taking into account the
Color versions of one or more of the figures in the article fact that the objective function could be nonlinear; this
can be found online at www.tandfonline.com/gcec. work also treated some of the variables as non-Gaussian.
218 U. Abubakar et al.

Grossmann et al. (1983) and You and Grossmann (2008) governing function(s) could be assumed to be linear,
presented an approach for planning and optimizing supply nonlinear, and either continuous or discrete.
chain design under uncertainty. In this work, probabilistic Grossmann et al. (1983) and Chacon-Mondragon and
models were used to predict stock levels, stock-outs, and Himmelblau (1988) presented a mathematical formulation
other demand uncertainties. of uncertain chemical engineering systems along the follow-
More recently, Moon et al. (2011) presented a framework ing lines. The set of parameters governing the performance
that could be used for integrating process design and control behavior of a chemical process is usually categorized into
under uncertainty. The study discussed a mathematical state variables ðwÞ, equipment sizes=dimension specifications
approach for reducing the combinatorial complexity ðdÞ, control variables ðcÞ, and vector of uncertain parameters
encountered in integrating design and control under uncer- ðxÞ. For a time-invariant process under uncertainty, a
tain parameters. Lu et al. (2010) applied a fuzzy modeling general mathematical formulation governing design and
approach to address similar problems. Sheen and Wang performance for a single-objective programming can be
(2011) also proposed an uncertainty quantification technique represented by (Edgar et al., 2001):
that employs polynomial chaos expansions (PCEs); using the
H2=CO=C1  C4 kinetic model and ethylene combustion  
data, uncertainties were quantified in the presence of some Minimize f d; c; w; x
   
constraints. The Advanced Process Engineering
subject to: h d; c; w; x ¼ 0; g d; c; w; x  0 ð1Þ
CO-Simulator (APECS), developed by National Energy
Technology Laboratory (NETL) of the U.S. Department
of Energy and other R&D collaborators, is an innovative where f() is the objective function, typically expressed in
software tool that combines steady-state process simulations monetary units, and h is the vector of design=performance
with computational fluid dynamics (CFD) and equations. The vector h could comprise mechanistic models
reduced-order models for optimization of overall process (obtainable from known principles such as mass, energy, and
performance (Zitney, 2010). Using the process industry stan- equilibrium equations) and=or regression models (obtained
dard, CAPE-OPEN (CO), APECS provides stochastic through response surface methodologies). The inequalities
analysis opportunity. Subramanyan et al. (2011) demon- vector, g, is often used to specify operating limits and pro-
strated the usefulness of this capability in the uncertainty duct quantity and quality, guided by, for instance, the cost
analysis of an advanced electricity and hydrogen generation of raw materials, utilities, market forces, and need for
facility. human, environmental, and equipment safety or certain
Some of the benefits and the need for stochastic methods regulatory requirements. From Equation (1), it can be seen
are demonstrated by the studies referenced above. However, that the vector of the state variables w is a subset of the other
the studies also demonstrate that many challenges to the suc- decision variables at a given design d for any realizations of
cessful application of stochastic methods remain and that x, suggesting that w can be expressed as an implicit function
opportunities exist to make use of methods developed in of the control variables c as:
other fields of engineering. This article presents a novel  
stochastic technique for designing and analyzing chemical h d; c; w; x; ¼ 0 ) w ¼ wðd; c; xÞ ð2Þ
engineering systems that contain uncertain variables. The
proposed approach combines response surface modeling
Resulting in a reduced inequality constraint:
(RSM) and structural reliability methods to model and
investigate the behavior of process systems under uncer-  
tainty. The article shows how commercial RSM software g d; c; wðd; c; xÞ; x ¼ G ðd; c; xÞ  0 ð3Þ
packages can be used in conjunction with chemical process
simulators and structural reliability analysis packages to Consequently, optimum design=performance realizations
model uncertainty in chemical engineering systems. Three are required (obtainable
  G ðd; c; xÞ  0) that minimize
from
case studies representing typical chemical engineering pro-
the cost function f d; c; w; x . Each of the uncertain para-
cesses are considered to validate the proposed approach.
meters ðxÞ is characterized by an associated range and a
given probability distribution. The resulting joint probability
Problem Formulation distribution is then integrated across the design=perfor-
mance space subject to G ðd; c; xÞ  0, which is set up to opti-
Process design and performance analysis under uncertainty mize the objective function. Numerical solution schemes can
involves two major tasks: formulation of functions that be employed to obtain probabilistic performance measures
describe performance and the adoption of solution techni- from the chemical process design=performance functions.
ques to solve the constrained objective function(s). The Some of the possible problems encountered in the course
mathematical formulations and solution strategies differ of this kind of analysis are: complexity or absence of objec-
essentially in the set of assumptions adopted relating to sys- tive function in a closed form, interactions between process
tem boundary definition, type of system (steady or transi- variables, non-normal tendency of random variables,
ent), parameter classification, operability ranges, parameter difficulty in input mapping=sampling, and the need for
uncertainty distribution, and variable dependence. The user-friendliness of the approach. These issues have received
A Hybrid Method for Modeling and Optimization 219

attention in the structural reliability research community dimensional nonlinearity. To address this problem, the
and will be of benefit to those in the chemical engineering results from response surface modeling are used to construct
community seeking to use similar probabilistic methods. In a simpler surrogate model.
this article, we present a novel application of such schemes Different mathematical functions can be fitted to the
applied widely in the field of structural reliability. The plant data or process simulation results to characterize their
first-order reliability method (FORM) has a number of performance, as discussed in Edgar et al. (2001). However,
strengths that would benefit chemical engineering design for the purpose of reliability analysis, polynomial functions
and analysis and will be highlighted in what follows. are usually fitted because of their ability to account for both
linear and nonlinear effects as well as interactions between
Methodology variables (Haldar and Mahadevan, 2000). This usually
requires nonlinear multiple regression analysis.
The approach proposed in this article is comprised of two If Y is an (N  1) vector of the responses, X is an (N  k)
major components: response surface modeling (RSM) and matrix of the input treatments, a is a (k  1) vector of the
stochastic analysis. The detailed steps include efficient regression coefficients, and e is an (N  1) vector of the ran-
response surface modeling using process simulation tools, dom errors, it can be shown that (Barton and Meckesheimer,
objective function development, solution of the objective 2006; Myers and Montgomery, 2002):
function using reliability techniques, and adjusting the con-
trol variables to achieve defined economic performance con- Y ¼ Xa þ e ð4Þ
strained by the need to ensure safety and reliability. Each is
described briefly in the following subsections. A common polynomial function linking the inputs and
response may then take the form:
Response Surface Modeling (RSM) XN XN1 XN XN
yðxÞ ¼ a0 þ a x þ
i¼0 1 1 i¼1
a xx þ
j<1 ij i j
a x2 þ e
i¼0 ii i
The aim of RSM is to establish a functional relationship
between the random inputs and the system response. After ð5Þ
defining the system to be modeled and the performance tar-
get to be achieved, the parameters are classified as either ran- where xi is the vector of the random input variables, N is the
dom or deterministic using available plant data or total number of random variables under consideration, a0,
experience of their effect on the outcome of the design prob- ai, aij, and aii are the regression coefficients to be calculated,
lem. The most probable sampling range for each of the and y(x) is the approximate objective function accounting
uncertain variables is then identified. for the fit errors. One benefit of the approach proposed in
Next, to ensure efficient treatment of the uncertain input this article is that the surrogate model needs to be accurate
variables, the sampling space is designed using one of the only at and around the region of high probability density,
experimental design schemes described in Myers and Mon- which is the main area of interest in reliability analysis (Hal-
tgomery (2002). There are a number of these design of dar and Mahadevan, 2000; Rajashekhar and Ellingwood,
experiment (DoE) techniques. Central composite design 1993). The total number of regression coefficients (Cr) to
(CCD) schemes are used to fit second-order response func- be determined depends on the number of selected variables
tions to data (Haldar and Mahadevan, 2000; Myers and N (Haldar and Mahadevan, 2000):
Montgomery, 2002). Popular examples of CCD schemes
are the Box-Wilson (BW) and Box-Behnken (BB) methods. ð N þ 1Þ ð N þ 2Þ
Cr ¼ ð6Þ
Choice of the experimental design method is governed by a 2
number of factors as discussed in Haldar and Mahadevan A number of commercial statistical software packages can
(2000), Liu and Moses (1994), and Najm (2009). A number be used to carry out the nonlinear multiple regression
of commercial software packages such as Design-Expert, analysis (Matlab was used in this study).
Matlab, and Statistica can be used to build the input
sampling; Design-Expert was used in this study.
Reliability Analysis
The process response to the random input variables is
then determined. There are two potential ways for achieving The aim of this reliability analysis is to determine the design
this: real experiments using existing plants or process simula- space and the corresponding region satisfying some pre-
tion software (for example, UniSim Design, Aspen HYSYS, scribed constraints, the most probable values of the
ProSimPlus, ChemCAD, COCO, gPROMS, or PRO=II). design=performance target, the corresponding values of the
The simulation tools are a familiar part of modern chemical input variables (design points, sensitivity of the target to
engineering and use encoded closed-form functions, i.e., the each of the random variables), chance of meeting=missing
mechanistic models obtained through mass=energy balances, the target, and the reliability=performance index. The results
thermodynamic principles, and reaction kinetics to model can be used to adjust the proposed design variables as
unit operations or processes along with their physical and appropriate to achieve the desired performance.
chemical properties. Even if available, physical models gov- With the objective function developed, the performance
erning the performance behavior of an entire process or even target (a threshold, limit state, or failure criteria) for the
a unit operation could be very complex, featuring high process is then defined, guided by experience, deterministic
220 U. Abubakar et al.

There are several techniques and algorithms for solving


Equation (8), including Monte Carlo simulation (MCS)
and its variants (such as directional simulation, importance
sampling, Latin hypercube sampling, Hammersley sampling,
and adaptive simulation) and first- and second-order
reliability methods (FORM=SORM) (Naess et al., 2012;
Thoft-Christensen and Baker, 1982). MCS is a well-known
technique, hence it will be only briefly highlighted here. On
the other hand, FORM analysis (Melchers, 1999), which is
relatively new, especially in chemical engineering applica-
tions, will be described in the next section. In order to esti-
mate the probability value using a simple MCS approach,
an indicator function or a counter I(x) is introduced to shift
the entire integration domain of Equation (8) to real space;
samples are then taken from the input vectors that are used
to solve the objective function. ‘‘1’’ is recorded whenever a
function response satisfies the prescribed constraint; other-
Fig. 1. Illustration of the performance space and the
wise, ‘‘0’’ is recorded. Matlab scripts may be coded to extract
Hasofer-Lind (H-L) reliability index b for a nonlinear safety
the responses. An estimate of the probability value is given
margin.
by the ratio of the sum of the successful responses to the
total number of responses. Note that the LSF is set up to
optimization results, available performance, failure or mar- constrain the objective function in order to realize certain
ket data, regulatory requirements such as the need to mini- design=performance targets. Extensive descriptions of MCS
mize the production of a certain pollutant, or other methods, including illustrative case studies, are given else-
factors. A mathematical function, GðX Þ, usually termed where (Labeau and Zio, 2002; Melchers, 1999; Naess et al.,
the limit state function (LSF), is then defined to split the 2012; Thoft-Christensen and Baker, 1982).
multidimensional state-space of the input variables into
regions representing ‘‘success’’ and ‘‘failure.’’ This surface First-Order Reliability Method
splits the domain as shown in Figure 1 (note that upper part
of Figure 1 could be the failure region depending on how the The first-order reliability method (FORM) seeks to deter-
LSF was defined). Considering a system under uncertainty, mine the probability of success or failure of the system being
we may for instance wish to determine the probability (Pf) analyzed by completing a numerical integration of Equation
that the system assumes a value, [y(x)], equal to or less than (8). In doing this, it determines the most probable design=
the threshold u. This can be expressed as: performance point ðx Þ of the input variables and a sensi-
tivity index for each variable that highlights their relative
importance and contribution to the calculated probability
Pf ¼ P½yðxÞ  u ¼ P½G ðX Þ ¼ yðxÞ  u  0 ¼ P½G ðX Þ  0 of failure. To avoid a number of problems linked to corre-
ð7Þ lation among the basic variables, the LSF is transformed
from physical (X) space to standard normal (Z) space,
which is a space of uncorrelated standard normal variables.
This transformation ensures that the point closest to the ori-
Also, in terms of the joint density function, the failure
gin of the failure domain ðx Þ is where the highest prob-
probability for the system may then be defined as
ability density portion is located and this is where failure is
(Thoft-Christensen and Baker, 1982):
most likely to occur. Putting this into context, the process
design=performance target is most sensitive to this point.
Z
: The transformation from physical to standard normal space
Pf ¼ GðX Þ0
fX ðxÞ  dx ð8Þ
ZN(0, 1) can be achieved using:

where GðX Þ is the limit state function, X is the input vector xi  lXi
zi ¼ ð9Þ
space, fX ðxÞ is the joint probability distribution function, rXi
and u is the defined threshold, which could be a constant
or a function. In this way, failure is deemed to have occurred Non-Gaussian variables can also be normalized through
whenever GðX Þ  0, and GðX Þ > 0 implies success or the normal tail transformation (Haukaas and Kiureghian,
non-failure. Note that, depending on the chosen threshold, 2006a):
design configuration, and specifications, the limit state func-
tion can be defined to accommodate the concept of process FXi ðxi Þ ¼ FZi ðzi Þ ¼ Uðzi Þ ð10Þ
flexibility, also called flexibility index, resilience index, or
process flexibility measure (Grossmann et al., 1983; Lasserre where FX i ðÞ is the cumulative distribution function of the
and Roubellat, 1985; Saboo et al., 1985). non-normal basic variable Xi and U() is the standard normal
A Hybrid Method for Modeling and Optimization 221

distribution function, hence: @G0


zi rG ðZi Þ @Zi
ai ¼ ¼ ¼  sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
  T ð17Þ
b jjrG ðZi Þjj 2
xi ¼ FX1i ½Uðzi Þ ð11Þ @G 0
@Zi

Similarly, with reference to the limit state surface, the cor-


responding coordinate of the point closest to the origin in Once the most probable design=operation point z is eval-
standard normal space is denoted by z . Geometrically, the uated, other performance measures such as the reliability
limit state surface can be visualized as an (n  1) dimensional index, probability of failure, and sensitivity indices can be
hyper-surface in the n  dimensional space of the basic vari- easily determined. Typical gradient search–based iterative
ables X (Thoft-Christensen and Baker, 1982). In the case of algorithms used to solve Equation (13) are: the Hasofer-Lind
nonlinear surfaces, the safety margin equation is linearized (H-L) algorithm for Gaussian variables (this was later
at the design point and replaced with a linear function extended to non-normal random variables by Rackwitz
(Koutsourelakis et al., 2004; Liu and Der Kiureghian, and Fiessler, leading to the HLRF algorithm), improved
1991); usually, a Taylor series expansion is employed for this HLRF (iHLRF) algorithm, gradient projection method,
purpose. Ignoring higher order terms, the linear function (a sequential quadratic programming, and Polak-He algorithm
hyper-plane for high dimensional systems) at the most prob- (Der Kiureghian et al., 2006; Liu and Der Kiureghian, 1991).
able design point ðz Þ in the standard normal space (i.e., the The basic steps involved in FORM analysis based on the
point with highest probability density) is given by: iHLRF algorithm, implemented in this study, are briefly
described below; more details can be found in Melchers
(1999) and Thoft-Christensen and Baker (1982).
G ðZ Þ  G ðz Þ þ rGðZi ÞT ðZi  zi Þ ð12Þ
Step 1: Based on the constructed objective function, a LSF

where z is the point on the failure surface that is closest to G ðX Þ is formulated and transformed to Zspace as
the origin. The length of the line segment between z and the explained earlier; the transformed LSF is then set to zero,
origin is denoted by b (as shown in Figure 1), termed the thus, G ðZ Þ ¼ 0.
Hasofer-Lind reliability index (Der Kiureghian et al., 2006; Step 2: The sensitivity indices, ai, with respect to each of the
Haukaas and Kiureghian, 2006b). So the problem has essen- basic variables are derived and evaluated using Equation
tially reduced to a constrained optimization problem, seek- (17). A general structure for the scheme used in searching
ing to minimize b along the surface G ðZÞ ¼ 0. It can be the design point is given as:
shown that:
Zmþ1 ¼ Zm þ Sm  D ð18Þ

 
z ¼ arg min jjzjjjG ðZ Þ ¼ 0 ð13Þ
where m is the iteration counter, D is the search direction,
where jjzjj is the norm of the vector z in the Euclidean space and Sm is the step size corresponding to the mth iteration.
representing the distance from the origin to the failure sur- The initial trial point is usually selected to be the mean
face in standard normal space. b can thus be expressed as: point on the limit state surface in physical space, which
corresponds to zero in the standard normal space. Both
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ai and G ðZ Þ are 0
XN
0 0
 evaluated at trial design point, z , to give
b¼ i¼1 i
z2 ; Subjected to GðZÞ ¼ 0 ð14Þ ai and G Z respectively.
Step 3: Based on iHLRF FORM algorithm, zi (given by
Equation (19)) is evaluated a number of times and in each
The probability of failure is given as:
case the corresponding value of the LSF is computed until
convergence is achieved, i.e., until G ðZ Þ ¼ 0 or r.
Pf ffi UðbÞ ð15Þ
 T
where U() is the standard normal cumulative distribution a0j z0j  Gðz0 Þ
function. Equation (15) gives an exact solution if the LSF zi ¼ a0i   T ð19Þ
is linear. Expressing Equation (12) in terms of the direction a0j
2

cosines (a) and the reliability index (b), it can be shown that
(Haukaas and Der Kiureghian, 2006a):
Step 4: The resulting coordinate, z , is used to calculate the
  required minimum distance b; see Equation (14). Some of
GðZ Þ  jjrGðz Þjj b  aT Zi ð16Þ the mathematical expressions given earlier can be
employed to compute other performance measures
where ai are the sensitivity indices, i.e., the direction cosines depending on the level of details required. Note that
of b along the respective coordinates evaluated at z , and Equation (9) can be used to map the design point z back
give quantitative measures of the sensitivity of the system to the physical space to gain x . The values can then be
performance at the design point to excitations in the basic used by the engineer to minimize the possibility of failure
variables; this can be given as: of the design.
222 U. Abubakar et al.

A number of software packages are available to undertake RSM approach described earlier, it can be seen that only cer-
FORM analysis, e.g.: FERUM 4.1 and OpenSees (both tain key parameters are considered initially (fostering dimen-
available free), and commercial software such as PROBAN, sional and correlation effects reduction), then the most
VaP, CalREL, and FeET. In this study, Finite Element probable operability range is identified from literature,
Reliability Using Matlab (FERUM) 4.1, developed by experience, or other sources. An experimental design
Bourinet et al. (2009), was used to calculate the performance strategy is then chosen to concentrate the sampling around
measures. FERUM was initially developed to handle struc- certain regions of interest in the operable range; this is very
tural reliability problems (Der Kiureghian et al., 2006), important as it gives the analyst leverage to direct the analy-
and it provides a wide range of operational options. Many sis to predetermined domains, saving computational time
of these options can be implemented for chemical process that would otherwise be used in finding the global extrema.
analyses depending on the problem formulation strategy The above description is visualized in the flowchart in
adopted. Figure 2.
The effort associated with FORM can be mitigated by With the sensitivity indices, reliability index, probability
how the RSM is performed. RSM can reduce the complexity of failure, and other statistics computed, the analyst can then
of the response function by reducing the dimensionality of adjust the specifications of the parameters to achieve the
the problem and promoting analysis efficiency, provided desired design=performance target.
the surrogate model is used within its intended limits. With
the introduction of linearization about the design point,
one of the major challenges has been addressed. Other prob- Case Studies Performed
lems include exhibition of more than one extremum (minima Three typical chemical engineering case studies were selected
or maxima) by chemical processes, pronounced curvature to apply the proposed hybrid stochastic framework. The first
around the design point, multidimensionality, and high cor- is plug flow reactor design and the second is heat exchanger
relation effects between the variables. Combining RSM with performance under uncertainty. Each of these two systems
the reliability techniques helps to address these issues. In the has a known design=performance function (in explicit form)

Fig. 2. A hybrid method for reliability-based performance modeling and optimization.


A Hybrid Method for Modeling and Optimization 223

and is assumed to be governed by three random variables. In The reaction rate is given by:
each case, two LSFs are developed, one based on the original
physical model and the other based on the surrogate model.
rA ¼ kCA ð22Þ
Reliability analysis is then carried out based on the two
models and the results are compared. The third case study The flow rate of phosphine and its conversion were con-
features a simple pump system considered by Swaney and sidered to be random. The reaction rate ‘‘constant’’ would
Grossmann (1985) and demonstrates additional benefits of be expected to vary due to changes in the concentration of
the proposed framework. phosphine along the length of the reactor, therefore it is
treated as a random variable. All three random variables
Case Studies are assumed to follow a Gaussian distribution. This is con-
trary to the approach adopted in Levenspiel (1999) for solv-
The main aim of choosing the three case studies is to serve as ing this particular problem where the randomness in F, k,
a means for highlighting the properties of the proposed and XA was not considered. The remaining variables are
stochastic analysis technique. The problems considered are assumed to be fixed, therefore handled as deterministic para-
simple and idealized to avoid complexity and show the meters. Operating details for the reactor are presented in
properties of the developed method. Uncertainties stemming Table I.
from the structure of the physical models are ignored in the Under the effect of uncertainty, we need to establish a
examples (i.e., process simulation results taken at face basis for an optimized design by propagating parametric
value). The demonstrative systems were sensitive to slight uncertainties, determining overall volume requirement and
changes in the decimal places, so longer decimal places were the sensitivity of the reactor size to the uncertain para-
deliberately adopted for some of the parameters (as shown in meters. Knowing the mean volume and the operability
some of the tables and RSM-based functions). Each of the ranges for each of the random parameters, we also wish
case studies is described and the results gained from to assess the plausibility of the design at a reduced volume
the RSM-FORM method are set out in the following (say 135 L instead of the mean 148L obtained through
subsections. deterministic calculations) by exploring the operability
ranges of the parameters (RF ¼ 40
14 mol=h, Rk ¼
Plug Flow Reactor Design 10
1.5 h1, and RXA ¼ 0:80
0:06Þ. The probability of
system failure and the performance=reliability index need
In a homogeneous gas process, phosphine is to be decom- to be determined.
posed into phosphorus and hydrogen. The decomposition
is intended to take place in a plug flow reactor defined by Solution
eight parameters namely, F: feed flow rate, (mol=h), k: reac- Through a simple material balance, a functional relationship
tion rate (=h), CA0: initial concentration of the phosphine between the volume and the design parameters can be
(mol=m3), P: pressure (kPa), T: temperature (K), XA: developed. First, the reactor is assumed to be idealized with
conversion, eA: expansion factor; and R: gas constant steady flow and negligible concentration gradient along the
(Pa  m3=mol  K). The problem was modified from Leven- radial directions. It is assumed that the concentration of
spiel (1999) where it was addressed as a deterministic case phosphine varies along the length of the reactor. Therefore,
with all the parameters assumed to be completely known to compute the volume of the reactor, there is a need to inte-
and fixed: grate a mechanistic model across the tubular reactor. For
this particular case, it can be shown (Levenspiel, 1999) that
4PH3 ðgÞ ! P4 ðgÞ þ 6H2 ðgÞ ð20Þ the volume is given by:

For brevity, the above equation is represented as: 


F 1
V¼ ð1 þ eA Þ ln  eA XA ð23Þ
4AðgÞ ! R4 ðgÞ þ 6S2 ðgÞ ð21Þ kCA0 1  XA

Table I. Design parameters for a plug flow reactor

Parameter Mean St. Dev Distribution

F, feed flow rate (mol=h) 40 14.00 Random (Gaussian)


k, reaction rate (=h) 10 1.50 Random (Gaussian)
XA, conversion 0.8 0.06 Random (Gaussian)
CA0, initial concentration (mol=m3) To be calculated 0.00 Deterministic
eA, expansion factor To be calculated 0.00 Deterministic
P, pressure (kPa) 460 0.00 Deterministic
T, temperature (K) 922 0.00 Deterministic
R, gas constant (Pa  m3=mol  K) 8.314 0.00 Deterministic
224 U. Abubakar et al.

Table II. Plug flow reactor performance GðX Þ ¼ V  0:135 ð27Þ

Reactor volume V (m3) Thus:


Input space from two functions 
F 1
Run F k XA Original RSM GOrig ¼ ð1 þ eA Þ ln  eA XA  0:135 ð28Þ
kCA0 1  XA
1 26.00 08.50 0.80 0.112999 0.113043
2 40.00 10.00 0.80 0.147768 0.147757 GRSM ¼ þ0:68440  3:42633F  103 þ 0:01409k  1:90234XA
3 40.00 10.00 0.80 0.147768 0.147757
 3:77923Fk  104 þ 0:01380FXA  0:05646kXA
4 40.00 11.50 0.74 0.104486 0.105278
5 40.00 11.50 0.86 0.162069 0.162098  3:23041F 2  107 þ 1:53983k2  103
6 54.00 11.50 0.80 0.173466 0.173400
þ 1:54577XA2  0:13500 ð29Þ
7 40.00 10.00 0.80 0.147768 0.147757
8 40.00 08.50 0.74 0.141363 0.141312 The equation for GRSM can be simplified by removing small
9 54.00 10.00 0.74 0.162214 0.161454 corrective terms (such as 3.23041 F2  107) that do not
10 40.00 08.50 0.86 0.219270 0.218455
affect the overall result. However, such terms are retained
11 40.00 10.00 0.80 0.147768 0.147757
for demonstrative purposes in this article. Also, note
12 26.00 11.50 0.80 0.083521 0.082721
13 54.00 08.50 0.80 0.234690 0.235467
that satisfying the LSF, G ðX Þ  0, means success in this
14 40.00 10.00 0.80 0.147768 0.147757 case (not failure) since the target here is reactor size
15 26.00 10.00 0.86 0.121147 0.121885 minimization.
16 54.00 10.00 0.86 0.251613 0.251613 Monte Carlo Simulation (MCS) Results
17 26.00 10.00 0.74 0.078103 0.078080 Taking into account the uncertainty in some of the variables
and the requirement for some of the parameters to be
assumed constant, MCS was used to generate performance
The initial phosphine concentration is given as: space for the reactor. In this case, the variables are assumed
P 460000 to be Gaussian for illustrative purposes. Figure 3 depicts a
CA0 ¼ ¼ ¼ 60 mol=m3 ð24Þ possible design space for the reactor showing interactions
RT 8:34  922
between F, k, and XA, while other variables are fixed. Note
The expansion factor is: that the space is designed ensuring that each of the uncertain
variables Vi satisfy a prescribed requirement such as
eA ¼ ð7  4Þ=4 ¼ 0:75 ð25Þ ri Vimax  Vi  Vimax for ri < 1, and subject to certain prob-
ability density function. Note that normal tail transform-
Adopting the Box-Behnken DoE scheme, the operability ation (see Equations (10) and (11)) may be needed in the
ranges for the three random variables are used to construct case of non-Gaussian variables.
the sampling space shown in Table II. This scheme is Having satisfied the above constraints, we further need to
adopted because of its ability to emphasize central points determine part of the performance space that guarantees
and the neighboring region, which is where considerable achievement of the prescribed design target, GðX Þ  0, in
interest lies in this case. The surrogate model for Equation this case P½V  0:135  0, i.e., in the presence of uncertain
(23) can be obtained as: variables, we need to explore the possibility of designing the
reactor at a volume equal to or less than 0.135 m3. Figure 4
VRSM ¼ þ0:68440  3:42633F  103 þ 0:01409k  1:90234XA
 3:77923Fk  104 þ 0:01380FXA  0:05646kXA
 3:23041F 2  107 þ 1:53983k2  103
þ 1:54577XA2 ð26Þ

Note that only three out of the eight parameters appear


actively in the surrogate model. The remaining five
deterministic parameters have been embedded in the model
and are accounted for in the reactor responses, provided that
the system remains within the given boundary conditions.
Reliability Analysis (Using FERUM 4.1)
Next, LSFs are set up based on the two models to assess
their degree of agreement. The intention is to compute: Fig. 3. Overall design space for a plug flow reactor depicting
interactions between F  N(40, 14), k  N(10, 1.5) and
P½V  0:135 ¼ P½V  0:135  0 ¼ P½GðX Þ  0 XA  N(0.8, 0.6).
A Hybrid Method for Modeling and Optimization 225
Table III. Sample MCS results for the plug flow reactor

Results based on:

Original limit state function RSM (Surrogate Model)

HL-index,

b¼ 0.195
HL-index, b ¼ 0.191
P GX ðxÞ  0 ¼ 0:4226 P GX ðxÞ  0 ¼ 0:4243

Table IV. These results give the operating conditions to


achieve about 80% phosphine conversion given that the reac-
tor volume must be at most 135 lL. The result, however, sug-
gests that there is only about 40% chance of meeting this
Fig. 4. Feasible space for the design of a plug flow reactor performance target, which obviously suggests that the design
depicting interactions between F  N(40, 14), k  N(10, 1.5) specifications need to be reconsidered, reliability needs to be
and XA  N(0.8, 0.6). stepped up by exploring the prescribed boundary conditions,
and FORM analysis may also be used to iteratively adjust
shows the space containing the set of design points that the specifications in order to achieve higher probability of
satisfy the LSF; in other words, this is a feasible design success. For instance, the alpha values quantify the sensi-
region for the plug flow reactor. Note that the space pre- tivity of the reactor size to the respective variables. The nega-
sented in Figure 4 is a subset of that shown in Figure 3. tive sign suggests that reducing the flow (F) and required
The two spaces are further contrasted on the same axes for conversion (XA) decreases probability of failure. However,
brevity; combined plots for the two scenarios are shown in phosphine conversion being essential in this case, reducing
Figure 5, showing the relative chance of achieving the flow rate may be a less expensive decision to take. On the
intended design objective; Table III shows the numerical other hand, a positive alpha value associated with the rate
values of P½G ðX Þ  0 and b. A hundred thousand MCS constant (k) suggests that the more such parameter is
runs were generated, and variations with increasing simula- increased the better the performance=design. Furthermore,
tion runs were noted, see Figure 6. Sample results based the magnitude of each of the indices can be used to rank
on the original reactor design model and the RSM-based the variables, the bigger the absolute value of the index,
functions are shown in Table III. It may be seen that both the bigger the impact of the corresponding variable; so in
point values (as tabulated) and the graphical values featur- this case, the system is most sensitive to F, then XA, and
ing about 10000 data points are in good agreement, indicat- finally k, These options are left to the decision maker to
ing that P½G ðX Þ  0 ffi 0:42. explore further.
But we still need to answer some questions like: What is
the most probable design point? What is the sensitivity of
Investigating the Uncertainty in the Heat Duty of a Heat
the reactor size to each of the basic variables? What is the
Exchanger Based on Experimental Results
reliability=chance of a successful reactor operation?
Also, we may be constrained by the need to make sure that, In the second case study, the performance of a
for instance: RF ¼ 40
14 mol=h, Rk ¼ 10
1.5 h1, and laboratory-scale counter current shell and tube Armfield
XA 80%. Note that a design setting to ensure no less than heat exchanger (Model HT 33) under uncertain input vari-
80% phosphine conversion is being sought, subject to the ables was investigated. The heat exchanger was used in con-
other constraints. To answer these questions, a FORM junction with a HT30X service unit. Using a PC, sensors,
iterative algorithm was used; sample results are given in and the associated software (the driver), operating

Fig. 5. Feasible and infeasible design spaces as predicted by the original and RSM based functions for F  N(40, 14), k  N(10, 1.5)
and XA  N(0.8, 0.6).
226 U. Abubakar et al.

Fig. 7. Random variation in the observed duty of a heat


Fig. 6. Variation of probability of successful reactor perfor- exchanger.
mance at volume, V  0.135 m3 with number of simulations.

G ðX Þ ¼ 1000  Q ð30Þ
conditions for the heat exchanger were sampled after every
10 s, resulting in a total of 330 experimental data sets. Based Original function:
on each sample set, the corresponding log mean temperature
difference (LMTD), overall heat transfer coefficient (U), GOrig ¼ 1000  4182:53F DTq ð31Þ
density (q), specific heat capacity (Cp), heat duty (Q), and
apparent heat transfer area (A) were calculated, taking into Surrogate model:
account the effect of temperature changes in each case. By
way of illustration, Figure 7 is presented to depict the 
GRSM ¼ 1000:00  939:542  2:72470F  107
observed random variation in the thermal duty of the heat
exchanger under review. As an excerpt from the experiments  1:43958DT  102  0:94126q þ 4:17599F DT  106
performed, data relevant to this study are presented in þ 2:72893F q  104 þ 0:14418qDT  0:05330F 2
Table V. 
 8:77000DT 2  108 þ 2:61173q2  107 ð32Þ
Solution
As in the previous case, the input sampling space is con- where DT is the temperature change between the cold and
structed using the Box-Behnken method. the hot streams. The corresponding space-containing perfor-
mance coordinates satisfying the constraint Q 1000 J=s are
Reliability Analysis
given by the surface plots in Figure 8. The surfaces are
Performance of the heat exchanger was observed to be obtained based on both original and RSM-based functions
uncertain, which is to be expected since the input variables as in the previous case. Sample MCS results are given in
are also random. The intention now is to determine the Table VI. The most probable performance point, sensitivity
chance that the heat exchanger assumes a heat duty greater of the heat duty to the basic variables, the reliability index,
than 1000 J=s and the necessary performance specifications and the probability of meeting the intended performance tar-
to achieve that. The LSF in this case can be defined as: get are obtained using the FORM algorithm, and sample
results are presented in Table VII. As should be expected,
P½Q 1000 ¼ P½1000  Q  0 ¼ P½G ðX Þ  0 from the sensitivity indices, it can be seen that the

Table IV. Performance measures for plug flow reactor based on FORM analysis

FORM results based on:


Original limit state function RSM (surrogate model)


HL-index, b ¼ 0.202 HL-index,

b ¼ 0.19557
P GX ðxÞ  0 ¼ 0:4197 P GX ðxÞ  0 ¼ 0:4225

Parameter a Design value Parameter a Design value

F 0.81856 37.679 F 0.81471 37.7714


XA 0.47213 0.79425 XA 0.4771 0.7944
k 0.3272 10.0997 k 0.32954 10.0969
A Hybrid Method for Modeling and Optimization 227
Table V. Statistics from experiemtal data for the heat exchanger

Parameter Mean St. dev. Distribution

F, hot water flow rate (m3=s) 0.0000344729 0.0000010693 Random (Gaussian)


DT, temperature change ( C) 6.5250000000 1.0680000000 Random (Gaussian)
q, water density, (kg=m3) 998.45000000 0.4240000000 Random (Gaussian)
CP, heat capacity, (J=kg C) 4182.5300000 0.0000000000 Deterministic

Fig. 8. Feasible performance space for the heat exchanger as obtained from original and RSM based function for F  N(3.447  105,
1.069  106), DT  N(6.525, 1.068) and q  N(998.450, 0.4242).

temperature effect dominates, followed by flow rate; changes expected deviations of þ200 and 550 kPa. Thus, the lower
in density appear to be relatively insignificant in the and upper bounds are 250 and 1000 kPa respectively. The
temperature range being considered in this example. exit pressure must be equal to or less than 1000 kPa. The
nominal value of the flow rate M is 10 kg=s with deviation
of þ2 and 5 kg=s, and the lower and upper bounds are 5
Performance of a Pump under Uncertainty
and 12 kg=s respectively. Given the uncertainty in some of
In the final case study, the proposed framework is applied to the parameters, the associated most probable design point,
solve an optimization problem considered by Swaney and reliability index, and sensitivity of the pump performance
Grossmann (1985). The authors presented an approach for are determined, assuming that the variables are Gaussian.
optimizing the operation of a pump, paying attention to per- The corresponding boundary conditions are:
formance target achievement and cost minimization, but did
not factor in the need for operational reliability. Also, in M2
some cases, the performance function for a system may Pv ¼ P1 þ qH   kM 1:84 D5:16  P2
qCv
not be available or could be present but in an implicit form.
On the other hand, the decision maker also needs an MH
approach that gives a wide range of performance measures, W
0:5
as discussed earlier.
We apply the approach presented in this article to solve rCvMAX  Cv  CvMAX
the problem and answer some of the questions raised above.
The pump system is described in Figure 9. The exit pressure 250  P2  1000 ð33Þ
of a pump system P2, has a nominal value of 800 kPa with
where Cv is the valve coefficient, D is the pipe diameter in m,
H is the pump head (kJ=kg), and W is the driver power in
Table VI. Sample MCS result for the heat exchanger kW.
Results based on: Relevant Pump Performance Data
Pump efficiency g ¼ 0.50; liquid density q ¼ 1000 kg=m3;
Original limit state function RSM (surrogate model) control valve range r ¼ 0.05; pressure drop constant
 0:86
HL-index, b ¼ 0.394245 HL-index, b ¼ 0.397769 kp ¼ 9.101  106 (kPa) kg (m)5.16; inlet pressure
s
P½GX ðxÞ  0 ¼ 0:3476 P½GX ðxÞ  0 ¼ 0:3478
P1 ¼ 100 kPa; outlet pressure control tolerance Ep ¼ 20 kPa;
228 U. Abubakar et al.

Table VII. Performance measures for the heat exchanger based on FORM analysis

Results based on:

Original limit state function RSM (surrogate model)

HL-index, b ¼ 0.38677 HL-index, b ¼ 0.38676


P½GX ðxÞ  0 ¼ 0:349464 P½GX ðxÞ  0 ¼ 0:349465

Parameter a Design value Parameter a Design value

F 0.19686 3.4554  105 F 0.19686 3.45543  105


DT 0.98043 6.92998 DT 0.98043 6.92997
q 0.00270 9.9845  102 q 0.00270 9.9845  102

Note that in this approach, all the uncertain parameters


(W, H, Cv, and M) have been integrated into the objective
function. Only H, Cv, and M are physically present in the
original objective function.

Fig. 9. Pump and piping system. Exposing the Pump to Parameter Uncertainty
In practice, process variables don’t remain at constant
values; small random fluctuations or larger process upset
conditions cause changes. In order to reflect reality, the
annualized cost Ca ¼ ca1D þ ca2W0.86 þ ca3 MHg ; constants: parameters are assumed to display some deviation from
Ca1 ¼ 5.6  105$=m; Ca2 ¼ 482.84 $=kW0.86; ca3 ¼ 500 $=kW. the optimum design specification values. For the purposes
From the data given, it can be shown that: of this case study, these deviations are assumed to be small
to ensure that the system is operating very close to the three
M2 trial operating conditions denoted by X 1 ; X 2 , and X 3 ,
GORiG ¼ 100 þ 1000H   ð9:101M 1:84  106 Þ
1000Cv shown in Tables VIII and IX. Each of the variables W, H,
 ð0:07625:16 Þ  1000 and M were assumed to have a standard deviation of

0.05 units and Cv being very small,
0.03 units, and D is
assumed to be constant.
First, it is assumed that the response function is not avail-
At these levels of variance around each of the three oper-
able: three sets of deterministic design specifications are pre-
ating conditions, the sample results, shown in Table IX, sug-
sented in Table VIII. Based on the given data, the input
gest that performance reliability needs to be improved.
space is designed using the Box-Behnken algorithm, and
However, note that the performance constraint G ðX Þ  0
the responses are obtained from the mechanistic model
has always been satisfied in each case. The proposed
(alternatively, this can be obtained through real experi-
approach was used to further analyze the pump system to
ments). The data were regressed to obtain the following
identify better operating conditionS and associated
function:
reliability level. For illustrative purposes, three additional
operating conditions (X 4 ; X 5 , and X 6 ) were selected, as
GRSM ¼ 21083:2  266:842W  24044:2H þ 1907:53Cv shown in Table X. Note that the performance specifications
 974:542M  1:38016WH  1010 and reliability values for the six cases are different, as
expected.
þ 2:53090WCv  1013  1:61272WM  1012
The approach gives the same results as presented in
þ 8:24646HCv  1012  5:69547HM  1011 Swaney and Grossmann (1985) (greyed part of Table X)
þ 382:948Cv M þ 4:64154W 2 þ 9658:40H 2 with additional performance measures such as operational
reliability and sensitivity indices. Note that both valve coef-
 682:883Cv2  21:8501M 2  1000:00 ð34Þ ficient (Cv) and flow rate (M) are control variables in the
analysis. From the sensitivity indices, it can be seen that
the system is most sensitive to Cv, and, fortunately, the econ-
Table VIII. Pump performance statistics omic function is not a function of Cv, which presents good
opportunity for optimization of the pump system. This also
Specification W H Cv M suggests that disparity in the Cv values between solution
approaches is inconsequential as far as the economic optimi-
1 23.75 1.187 5.039000 5 zation is concerned; of course this implies idealization,
2 31.20 1.300 0.005770 8.5
aimed at simplifying the problem. Hence, this advantage
3 33.74 1.406 0.002969 12
was taken by adjusting the Cv specifications (the new values
A Hybrid Method for Modeling and Optimization 229
Table IX. Sample design points (X 1 ; X 2 ; and X 3 ) and the associated reliability indices

X 1 X 2 X 3

HL-index, b ¼ 0.0118 HL-index, b ¼ 0.0129 HL-index, b ¼ 0.00983


P½GX ðxÞ  0 ¼ 0:495 P½GX ðxÞ  0 ¼ 0:505 P½GX ðxÞ  0 ¼ 0:504

Parameter a Design value Parameter a Design value Parameter a Design value

Cv 0.699 4.210 Cv 0.691 2.434 Cv 0.431 2.929


H 0.669 1.187 H 0.674 1.300 H 0.896 1.406
M 0.253 5.000 M 0.259 8.500 M 0.108 12.000
W 0.028 23.750 W 0.014 31.200 W 0.013 33.740

Table X. Economic vs. reliability based performance optimization

Variable

Design specification W (kW) H (kJ=kg) Cv M (kg=s) Reliability

X 1 23.750 1.187 5.039000 (4.210000) 5.000 0.495


X 2 31.200 1.300 0.005770 (2.434000) 8.500 0.505
X 3 33.740 1.406 0.002969 (2.929000) 12.000 0.504
X 4 25.000 1.514 1.700 6.996 0.763
X 5 23.499 1.634 1.370 6.985 0.999
X 6 23.998 1.498 1.983 7.977 1.000

are parenthesized in Table X) resulting in more optimum point ðx Þ as obtained from FORM analysis. The index
design points without significant impact on the overall capi- can be stepped up (or down, depending on cost-benefit
tal investment, since the annualized cost of operating the analysis) by adjusting the specification of one or more of
pump is not a function of Cv, as mentioned earlier. Also, the basic variables based on the sensitivity indices.
note that each of the operating conditions in Table X satis- This method can be used to investigate processes with
fies the prescribed conditions given in Equation (1). The simple, complex, or no mechanistic model at all at both
decision maker is therefore left with three major the design and operational stages. To develop the response
performance measures on which decision can be based: function, both the input-space and responses need to be
performance target achievement, (GOrig  0), operational known; while an efficient experimental design provides the
reliability, and cost efficiency; sample decision options input sampling space, the responses can be generated in
(X 4 ; X 5 , and X 6 ) are presented in Table X. two major ways: from an existing explicit full-fidelity model
or a real experiment, giving some flexibility to the user.
Further Discussions Use of an efficient experimental design (strategized sam-
pling) technique such as CCD makes it possible to empha-
From the problem formulation and the solution strategies size specific regions of interest within the operability range
described, it can be seen that the term H-L reliability index of the variables, allowing for previous experience on the sys-
(b), in this context, is a probabilistic measure of the ability tem behavior to be incorporated into the analysis, and sav-
of a process to operate in a variety of conditions while meet- ing considerable effort that would otherwise be used in
ing the intended performance target. In a way, this index random sampling. Functions other than polynomials can
gauges the robustness, resilience, or flexibility of the facility be used if necessary. But polynomial functions appear to
in the face of random basic variables. The problem was for- be particularly good for reliability analysis due to their sim-
mulated (as discussed in the first section) and solved (see the plicity and ability to account for nonlinear effects and inter-
second section) in such a way that the magnitude of the actions between variables. It is also interesting to note that
reliability index increases as the probability of failure commonly found software packages (some from open
decreases. Note that what is good or bad probability value sources) can be combined to apply the proposed method.
is entirely dependent on the initial definition of ‘‘failure,’’ Also, even though the examples discussed earlier are gov-
which could be subjective. This is a very important decision erned by Gaussian random variables, in a similar way, the
an analyst has to make and affects the resulting performance framework can be easily used to handle non-Gaussian vari-
measures significantly. In addition, the approach characterizes ables, as indicated by Equation (11); see Melchers (1999) and
a process facility with respect to the design=performance Thoft-Christensen and Baker (1982) for more details.
230 U. Abubakar et al.

Finally, in using this framework it is important to pay CA0 initial concentration of phosphine
special attention to parameter characterization in terms of Cv valve coefficient
uncertainty range and distribution. It is also important to Cp specific heat capacity
note that the focus of the described approach is on d specifications equipment sizes=dimension
time-invariant processes. For transient processes, a time D pipe diameter
dimension needs to be added to the problem formulation. F feed flow rate of phosphine
GOrig original objective function
Conclusions GRSM surrogate objective function
G ðX Þ limit state function
This article presents an approach for optimizing the design H pump head
and performance of process systems under uncertainty. k reaction rate
Structural reliability techniques are combined with response kp pressure drop constant
surface methodology as well as process simulation to solve m iteration counter
constrained stochastic optimization problems. The frame- M nominal value of the pump flow rate
work provides a simple method of generating a wide range N(l, r) normal distribution with mean l and standard
of probabilistic decision indices that can support chemical deviation r
engineers as they seek to improve plant safety, reliability, P pressure
and financial performance. The use of DoE in constructing P1=P2 inlet=out pressure
response function saves considerable effort that would Pf probability value
otherwise be used in random sampling, often with dispro- Q heat duty
portional gain. It also allows the user to emphasize certain r control valve range
regions of interest within the parameter range; this way, Sm step size
experience on the parameters can also be incorporated into U overall heat transfer coefficient
the response function. Sample application case studies repre- W driver power
senting typical chemical engineering processes were con- x=X uncertain parameters
sidered to validate the proposed framework. The results XA conversion of phosphine
obtained support the usage of this framework in investigat- z most probable design=operating point
ing the performance of chemical processes under uncer-
tainty. Unlike traditional deterministic methods, stochastic Greek letters
modeling approaches, such as the framework proposed in ai sensitivity index for ith variable
this article, ensure deeper understanding of process perfor- b reliability index
mance behavior under uncertainty. Constrained by the need D search direction
to minimize cost and ensure safety and reliability, and DT temperature change
guided by the magnitude and directions (
) of the obtained e vector of the random errors
sensitivity indices, it is possible to optimize chemical engin- eA expansion factor
eering processes by implementing the proposed approach. Ep outlet pressure control tolerance
To achieve that, MCS and FORM analysis along with q density
RSM may be used to determine various stochastic perfor- g pump efficiency
mance measures, including the most probable design= (w) state variables
operation specifications, under given boundary conditions. u performance criteria=threshold

Funding References
Arellano-Garcia, H., and Wozny, G. (2009). Chance constrained opti-
The authors wish to gratefully acknowledge the financial mization of process systems under uncertainty: I. Strict monotoni-
support granted by Petroleum Technology Development city, Comput. Chem. Eng., 33(10), 1568–1583.
Fund (PTDF), Nigeria. Sriramula’s work within the Lloyd’s Ascough II, J. C., Maier, H. R., Ravalico, J. K., and Strudley, M. W.
Register Foundation Centre for Safety and Reliability (2008). Future research challenges for incorporation of uncertainty
Engineering at the University of Aberdeen is supported by in environmental and ecological decision-making, Ecol. Model.,
Lloyd’s Register Foundation (LRF). LRF, a UK registered 219(3–4), 383–399.
Badar, M. A., Zubair, S. M., and Sheikh, A. K. (1993). Uncertainty
charity and sole shareholder of Lloyd’s Register Group Ltd., analysis of heat-exchanger thermal designs using the Monte Carlo
invests in science, engineering, and technology for public simulation technique, Energy, 18(8), 859–866.
benefit, worldwide. The authors would also like to thank Barton, R. R., and Meckesheimer, M. (2006). Metamodel-Based
the anonymous reviewers for their constructive comments. Simulation Optimization, 535–574, Elsevier, Amsterdam.
Bourinet, J. M., Mattrand, C., and Dubourg, V. (2009). A review of
recent features and improvements added to FERUM software,
Nomenclature paper presented at the 10th International Conference on Structural
Safety and Reliability (ICOSSAR’09).
A apparent heat transfer area Chacon-Mondragon, L., and Himmelblau, D. M. (1988). A new
c control variables definition of flexibility for chemical process design, Comput. Chem.
Ca annualized cost Eng., 12(5), 383–387.
A Hybrid Method for Modeling and Optimization 231
Cheng, H., and Sandu, A. (2009). Uncertainty quantification and Naess, A., Leira, B. J., and Batsevych, O. (2012). Reliability
apportionment in air quality models using the polynomial chaos analysis of large structural systems, Prob. Eng. Mech., 28,
method, Environ. Model. Softw., 24(8), 917–925. 164–168.
Der Kiureghian, A., Haukaas, T., and Fujimura, K. (2006). Structural Najm, H. N. (2009). Uncertainty quantification and polynomial chaos
reliability software at the University of California, Berkeley, Struct. techniques in computational fluid dynamics, Annu. Rev. Fluid
Saf., 28(1–2), 44–67. Mech., 41, 35–42.
Diwekar, U. M., and Rubin, E. S. (1991). Stochastic modeling of Pistikopoulos, E. N., and Grossmann, I. E. (1988). Stochastic optimiza-
chemical processes, Comput. Chem. Eng., 15(2), 105–114. tion of flexibility in retrofit design of linear systems, Comput. Chem.
Edgar, T. F., Himmelblau, D. M., and Lasdon, L. S., eds. (2001). Eng., 12(12), 1215–1227.
Optimization of Chemical Processes, 2nd ed., McGraw Hill, New York. Pistikopoulos, E. N., and Mazzuchi, T. A. (1990). A novel flexibility
Grossmann, I. E., Halemane, K. P., and Swaney, R. E. (1983). analysis approach for processes with stochastic parameters, Comput.
Optimization strategies for flexible chemical processes, Comput. Chem. Eng., 14(9), 991–1000.
Chem. Eng., 7(4), 439–462. Rajashekhar, M. R., and Ellingwood, B. R. (1993). A new look
Haldar, A., and Mahadevan, S. (2000). Reliability Assessment Using at the response surface approach for reliability analysis, Struct. Saf.,
Stochastic Finite Element Analysis, John Wiley & Sons, New York. 12(3), 205–220.
Haukaas, T., and Kiureghian, A. D. (2006). Strategies for finding the Saboo, A. K., Morari, M., and Woodcock, D. C. (1985). Design of
design point in non-linear finite element reliability analysis, Prob. resilient processing plants—VIII. A resilience index for heat
Eng. Mech., 21(2), 133–147. exchanger networks, Chem. Eng. Sci., 40(8), 1553–1565.
Koutsourelakis, P. S., Pradlwarter, H. J., and Schuëller, G. I. (2004). Sheen, D. A., and Wang, H. (2011). The method of uncertainty
Reliability of structures in high dimensions, Part I: Algorithms quantification and minimization using polynomial chaos expan-
and applications, Prob. Eng. Mech., 19(4), 409–417. sions, Combust. Flame, 158(12), 2358–2374.
Labeau, P. E., and Zio, E. (2002). Procedures of Monte Carlo transport Subramanyan, K., Diwekar, U., and Zitney, S. E. (2011). Stochastic
simulation for applications in system engineering, Reliab. Eng. Syst. modeling and multi-objective optimization for the APECS system,
Saf., 77(3), 217–228. Comput. Chem. Eng., 35(12), 2667–2679.
Lasserre, J., and Roubellat, F. (1985). Measuring decision flexibility in Swaney, R. E., and Grossmann, I. E. (1985). An index for operational
production planning, IEEE Trans. Automat. Contr., 30(5), 447–452. flexibility in chemical process design. Part I: Formulation and
Levenspiel, O. (1999). Chemical Reaction Engineering, 3rd ed., John theory, AIChE J., 31(4), 621–630.
Wiley & Sons, New York. Szépvölgyi, Jr. J., Dı́az, E., and Gyenis, J. (1999). New stochastic
Liu, P., and Der Kiureghian, A. (1991). Optimization algorithms for modelling of mixing in process operations, Chem. Eng. Process.:
structural reliability, Struct. Saf., 9(3), 161–177. Process Intens., 38(1), 1–9.
Liu, Y. W., and Moses, F. (1994). A sequential response surface Thoft-Christensen, P., and Baker, M. J. (1982). Structural Reliability
method and its application in the reliability analysis of aircraft Theory and Its Applications, Springer-Verlag, New York.
structural systems, Struct. Saf., 16(1–2), 39–46. Vasquez, V. R., and Whiting, W. B. (2004). Analysis of random and
Lu, X., Li, H.-X., Duan, J.-A., and Sun, D. (2010). Integrated systematic error effects on uncertainty propagation in process
design and control under uncertainty: A fuzzy modeling approach, design and simulation using distribution tail characterization,
Ind. Eng. Chem. Res., 49(3), 1312–1324. Chem. Eng. Commun., 191(2), 278–301.
Melchers, R. E. (1999). Structural Reliability: Analysis and Prediction, You, F., and Grossmann, I. E. (2008). Design of responsive supply
2nd ed., Wiley, Chichester. chains under demand uncertainty, Comput. Chem. Eng., 32(12),
Moon, J., Kim, S., and Linninger, A. A. (2011). Integrated design and 3090–3111.
control under uncertainty: Embedded control optimization for Zio, E. (2009). Reliability engineering: Old problems and new
plantwide processes, Comput. Chem. Eng., 35(9), 1718–1724. challenges, Reliab. Eng. Syst. Saf., 94(2), 125–141.
Myers, R. H., and Montgomery, D. C., eds. (2002). Response Surface Zitney, S. E. (2010). Process=equipment co-simulation for design and
Methodology: Process and Product Optimization Using Designed analysis of advanced energy systems, Comput. Chem. Eng., 34(9),
Experiments, 2nd ed., John Wiley & Sons, New York. 1532–1542.

You might also like