Journal of Loss Prevention in the Process Industries 51 (2018) 112–124

Contents lists available at ScienceDirect

Journal of Loss Prevention in the Process Industries

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

Inherently safer design of a reactor network system: A case study T

a,∗ b a a
Saeed Eini , Majid Javidi , Hamid Reza Shahhosseini , Davood Rashtchian
a
Center for Process Design, Safety and Loss Prevention (CPSL), Chemical and Petroleum Engineering Department, Sharif University of Technology, Tehran, Iran
b
Iranian Offshore Engineering and Construction Company (IOEC), Iran

A B S T R A C T

The strategies of Inherently Safer Design (ISD) provide a conceptual approach in order to design equipment and
processes with substantially improved safety level. However, this may lead to a less economically attractive
design. This study aimed to obtain optimal decision parameters of a reactor network system to produce allyl
chloride. The objective functions were the risk level, including the severity and the frequency of the accidents,
which were associated with the hazards in the network and the economic profit of the process. Based on this
optimization approach, an array of optimal solutions (called Pareto front) was obtained as a trade-off between
the objectives under investigation. A final design point was ultimately selected using Shannon's entropy and
Bellman-Zadeh's techniques of decision making in a fuzzy environment. Results showed that the optimum re-
actor network leads to a highly complex system and more process control difficulties. This result was inconsistent
with the simplification strategy of Inherently Safer design. In order to deal with this problem, a sensitivity
analysis was performed that yielded a decision guide to decide about the desirable level of the risk as well as the
optimum design.

1. Introduction operations (Thiruvenkataswamy et al., 2016). Although it is applicable

Reactor systems usually represent a large portion of the total risk in
a chemical process. Innovative reactor designs that improve mixing of
the reactive fluids may result in much smaller reactors. Such designs are
usually cheaper to build and operate, as well as being safer due to their
smaller inventory. Consequently, safer design of rectors system, as the
main part of the processes, is crucial in order to make the overall
process safer. This can be done using any category of the risk reduction
strategies, including Inherent, Passive, Active, and Procedural (CCPS,
2009).
The two most important tools to make a plant safer are inherently
safer design philosophy in the preliminary stages of the design and the
operational acts to control hazards. A process that is inherently safer
will require fewer and less robust layers of protection. This is due the
fact that an inherently safer design can either reduce the magnitude of a
potential incident or make the occurrence of the accident highly un-
likely (Mannan, 2012).
There are four basic principles of inherently safer design. These
principles are shown in Table 1 (Shariff and Zaini, 2010).
Development of a chemical process consists of many sequential
stages, namely research and development (R&D), conceptual design,
detailed design, construction, start-up and commissioning, and plant
to implement inherently safer design concepts and to improve the in-
herent safety characteristics of a plant during any stage of its lifecycle,
the greatest opportunities for making major improvements in inherent
safety are provided in the first stage (R&D), when the designer may
have many choices of basic technology and chemistry available, and
may be free to choose less hazardous alternatives (Mannan, 2012).
Generally, the evaluation and comparison of the inherent safety
level with respect to different design options can be categorized in two
groups (Eini et al., 2015) as follows:
• Evaluation by scoring the process features and the development of
indices,
• Quantitative assessment using consequence modeling and the cal-
culation of accidents' consequences,
The majority of the attempts have been made in order to obtain
indices to evaluate the inherent safety level (Eini et al., 2015). Several
studies reviewed available tools and techniques for evaluating inherent
safety using indices (Khan and Amyotte, 2005; Kletz and Amyotte,
2010; Khan et al., 2015). Despite the simplicity of using inherent safety
indices, they only provided a relative evaluation of the level of risk
between different design options and did not consider vulnerable

∗
Corresponding author. Center for Process Design, Safety and Loss Prevention (CPSL), Department of Chemical and Petroleum Engineering, Sharif University of Technology, Azadi
Tehran, Iran.
E-mail address: eini_saeed@che.sharif.edu (S. Eini).

https://doi.org/10.1016/j.jlp.2017.12.004
Received 22 June 2017; Received in revised form 8 December 2017; Accepted 8 December 2017
Available online 09 December 2017
0950-4230/ © 2017 Elsevier Ltd. All rights reserved.
S. Eini et al.
Table 1
Pareto frontier of a multi-objective optimization.
Principle Definition
Intensification Also known as minimization. Use smaller quantities of
hazardous substances (also called intensification)
Substitute Replace a material with a less hazardous substance
Moderate Use less hazardous conditions, a less hazardous form of a
material, or facilities that minimize the impact of a release of
hazardous material or energy (also called attenuation and
limitation)
Simplify Design facilities which eliminate unnecessary complexity and
make operating errors less likely, and which are forgiving of
errors that are made (also called error tolerance)
elements in the surrounding environment as the possible hazard re-
ceptors. However, more important point is that these indices did not
demonstrate the possible, and often important, economic benefits of
implementing inherent safety. But, evaluating inherent safety using
consequence modeling approach considers hazard receptors and pro-
vides a clearer understanding of the risk (Eini et al., 2015). Table 2
shows the studies that have considered consequence modeling in order
to evaluate inherent safety level.
It should be noted that, it is almost impossible to simultaneously
maximize all desired characteristics of a process. The central problem of
all engineering designs is to find the optimum combination of char-
acteristics that best meets the overall objectives (Mannan, 2012). In-
herent safety is one of the many characteristics of a process, which
should be evaluated as an optimization problem. In many cases, making
a process inherently safer may conflict with the economics of the plant.
On the other side, reducing a hazard may result in increasing another
hazard. Consequently, for decision making and designing an “eco-
nomically optimum and inherently Safer” process, it is important to
consider a multi-objective optimization algorithm in which different
objective functions are addressed to reach their optimum values; one
objective function with respect to risk level and another one with re-
spect to economics. Multi-Objective Optimization (MOO) method is an
efficient approach for optimizing the problems dealing with conflicting
objectives. Table 3 represents the research studies that have considered
multi-objective formulation for inherently safer design.
Although, Medina-Herrera et al. (2014) proposed a multi-objective
formulation and performed a sensitivity analysis in their study, they
didn't solve the problem using MOO.
In order to assess a realistic measure of risk, it is crucial to consider
the frequency of the accidents as well as the consequences. Eini et al.
(2015) did not employ the probabilistic approach in risk assessment.
They considered economic damages arising from the accidents as the
measure of inherent safety and combined the objective functions in
order to convert the multi objective problem into a single objective.
Moreover, Eini et al. (2016a) conducted the same case study as the
previous research. However, they combined detailed quantitative risk
assessments considering the frequencies and consequences of the in-
cidents. Indeed, they utilized NSGA-II in order to solve the MOO for-
mulation. In order to develop MOO framework for inherently safer
design, Eini et al. (2016b) considered exergy efficiency as a third ob-
jective besides the economic and safety objectives for a cascade re-
frigeration unit.
To the best of the authors' knowledge, there is no comprehensive
study on the synthesis of inherently safer reactor networks in the open
literature. In this paper, in order to design an inherently safer reactor
network for allyl chloride production, a multi-objective optimization
(MOO) of the system was performed in order to minimize the risk posed
by the hazards in the network and maximize the process of the eco-
nomic profit as objective functions. In the allyl chloride process, both
toxic and flammable hazards are expected. This suggests an opportunity
for the implementation of ISD's strategies in order to make the process
113
Journal of Loss Prevention in the Process Industries 51 (2018) 112–124
inherently safer. In the analysis, both Intensification and Moderation
strategies were considered. In this paper, the capital and maintenance
costs of system components, the operational cost, and the profit due to
selling of the product were included in the process of the economic
profit of the plant. In order to calculate the frequency of the probable
accidents, a multivariable model based on event tree was used, which
yielded more accurate risk calculation. In this study, an evolutionary
algorithm (EA) method based on the elitist NSGA-II was employed to
obtain the Pareto optimal set and Pareto frontier in objective space. In
the next step, Euclidian and fuzzy non-dimensionalization was em-
ployed for non-dimensionalization of the Pareto front results, and the
final optimal solutions from available solutions located at the Pareto
frontier were selected by the decision makers, including Shannon
Entropy and Fuzzy methods proposed by Bellman-Zadeh.
2. Economically optimum-inherently safer design of a reactor
network system
Economic profitability is the main goal of the design of a reactor
system. However, several additional considerations such as safety as-
pects, environmental regulations, etc. should be respected in the design
work. These considerations may have either direct or indirect effects on
overall plant profits. However, safety aspects, has the most important
effects on the overall profits. On the one hand, this issue arises from the
fact that controlling hazards in a plant requires safety functions that
need a considerable efforts and costs in chemical processes. On the
other hand, occurrence of an accident produces all types of direct and
indirect costs (Reniers and Audenaert, 2009). Therefore, if a hazard
leads to an accident, large monetary losses would be imposed to the
plant. Accordingly, one of the aspects that should be considered in plant
design is process safety to reduce the overall risk level of the plant. This
can be done using any category of the risk reduction strategies, in-
cluding Inherent, Passive, Active, and Procedural (CCPS, 2009). In-
herent safety principles not only can reduce or maybe eliminate the
high risk nature of the plant (both accidents magnitude and probability
of occurrence), but also reduce the required expenses in safety func-
tions.
It should be noted that different process options and alternatives
differ in level of potential of accident occurrence and severity. It is
possible that a design, which has lower processing costs, has more as-
sociated hazards and consequently more accident costs (Eini et al.,
2015). To perform a realistic optimization, all costs associated with a
process, including accident costs, which are influenced by decision
parameters of a design procedure, should be taken into account. Con-
sequently, it is essential to find an “Economically Optimum-Inherently
Safer design” using an optimization framework.
In order to deal with risk level as a measure of inherent safety,
probabilistic risk assessment can be performed in which the con-
sequence and the frequency of the probable accidents are combined.
Therefore, any effort to lower the calculated risk leads to higher level of
inherent safety. This risk reduction should be done simultaneously in
connection with enhancing the profits of the plant using a Multi-
Objective Optimization (MOO) procedure. This procedure is briefly
described below and is presented in more details in Section 3.
2.1. Objective functions: the overall risk level and profit of the plant
There are many alternatives, options, and operational conditions
that should be determined to finalize a process design. For reactor
network systems, the following options may be considered:
✓ Different reaction routes to obtain a specific product,
✓ Contacting pattern and technology (plug flow reactor, mixed flow
reactor, etc.),
✓ Number of reactors in series or parallel,
✓ Operation condition (temperature distribution along the reactor,
S. Eini et al. Journal of Loss Prevention in the Process Industries 51 (2018) 112–124

Table 2
Studies on the inherent safety based on the consequence modeling.

Reference Reference's details

Shariff et al. (2006) Subject Evaluation the consequence of VCE to assess the level of inherent safety among different designs at initial simulation
stages
Approach Developing a demonstrative tool called “integrated Risk Estimation Tool” (iRET).
Tools Process simulation using HYSYS
Consequence modeling developed in MS Excel
Case study Piping leak from liquefied natural gas (LNG) plant
The effect of introducing the inert (steam) on explosion risk acrylic acid production
Determining preliminary layout of the plant concurrently with process design.
Considered Hazards VCE
Considered Strategies Moderation
Shariff and Zaini (2010) Subject development of a prototype tool for consequence analysis and design improvement via inherent safety principle
Approach Developing a tool called “Toxic release consequence analysis tool” (TORCAT)
Tools The consequence analysis based on the worst-case scenarios
Case study identify potential hazards from selected streams for a purification column in a typical ammonia production plant
Considered Hazards Toxic Release
Considered Strategies Moderation
Intensification
Patel et al. (2010) Subject Inherently safer design of solvent processes at the conceptual
Stage
Approach Integrating consequence modeling into process simulation to evaluate the safe distance as a measure of inherent safety for
different design alternatives
Tools Aspen plus simulator
ICAS 11.0-ProCAMD (Computer Aided Molecular Design software) consequence modeling (Using regulatory guidance
from EPA RMP)
Case study Liquid extraction of acetic acid-water mixture
Considered Hazards VCE
Considered Strategies Intensification
Moderation
Substitution
Medina-Herrera et al. (2014) Subject Development of inherently safer distillation systems
Approach using principles of quantitative risk analysis together with economic objectives
Tools ASPEN Plus process simulator
A toolkit developed in MATLAB to calculate societal and individual risk and cost
Case study This case study considers a mixture of hydrocarbons, n-hexane/n-heptane, to be separated by distillation
Benzene-toluene mixture that must be separated by distillation
Considered Hazards Jet fire
Flash Fire
BLEVE
VCE
Flash Fire
Toxic release
Considered Strategies Intensification
Moderation
Simplification
Eini et al. (2015) Subject Developing an optimization procedure to select inherently safer design Scheme
Approach Optimization of an objective function that includes processing costs and accident costs
Tools Aspen HYSYS Process simulator
A toolkit (integrated with simulator) developed in MATLAB to model the consequences and evaluate processing costs
Case study Simple refrigeration cycle
Considered Hazards BLEVE/Fireball
Considered Strategies Intensification
Substitution
Moderation
Eini et al. (2016a) Subject Multi-Objective formulation in order to inherently safer design of a simple refrigeration cycle
Approach Optimization of total annual cost and risk level simultaneously
Tools Aspen HYSYS Process simulator
A toolkit (integrated with simulator) developed in MATLAB in order to model the consequences, assess accident
frequencies, and evaluate total annual cost
Case study Simple refrigeration cycle
Considered Hazards BLEVE/Fireball
Considered Strategies Intensification
Moderation
Eini et al. (2016b) Subject Multi-Objective formulation in order to inherently safer design of a simple refrigeration cycle
Approach Optimization of total cost rate (considering environmental penalty), exergy efficiency, and risk level simultaneously
Tools Aspen HYSYS Process simulator
A toolkit (integrated with simulator) developed in MATLAB in order to evaluate risk level, total annual cost,
environmental damages, and exergy efficiency
Case study Cascade refrigeration cycle
Considered Hazards Toxicity/BLEVEa/VCEb/VCFc/Jet Fire
Considered Strategies Intensification
Substitution
Moderation

a
Boiling Liquid Expanding Vapor Explosion
b
Vapor Cloud Explosion

114
S. Eini et al. Journal of Loss Prevention in the Process Industries 51 (2018) 112–124

c
Vapor Cloud Fire

Table 3 Non-dominated soluƟon

Studies that have considered multi-objective formulation for inherently safer design.

Reference Objective Functions Optimization algorithms

Medina-Herrera et al. • Total annual cost – 1

(2014) • Product purity Ideal soluƟon
• Societal risk Dominated
Eini et al. (2015) • Processing cost Direct search
• Accident cost
soluƟon
Eini et al. (2016a) • Total annual cost NSGA-II a

• Risk level
Eini et al. (2016b) • Total cost rate NSGA-II a

• Total exergy efficiency

• Total risk level
a
Non-dominated Sorting Genetic Algorithm II.
pressure),
✓ Intermediate separation of products,
✓ Bypass and recycle streams,
✓ Using inert as a dilution agent to control reaction temperature and
rate,
✓ Using reactive distillation to combine reaction and separation steps,
These design variables may be selected according to inherent safety
principles. All of the possible options should be embedded in a reactor
network superstructure to consider the entire possible design scheme.
Fixing the values of different design variables in the superstructure
leads to a possible scheme that should be further evaluated to assess the
level of inherent safety and the economic attractiveness.
As was previously mentioned, to evaluate the level of inherent
safety, a probabilistic risk assessment can be performed. For a specified
design scheme, hazardous scenarios should be identified. For these
scenarios, Fault Tree Analysis (FTA) and Event Tree Analysis (ETA)
should be developed to determine scenario outcome accidents and their
occurrence frequencies. Then, it is essential to evaluate the severity of
the identified probable accidents using consequence modeling. The
consequences may be expressed in terms of the affected area, affected
people, a specific damage level, or monetary losses. Risk level, as the
first objective function, can be calculated combining the consequence of
any accident with its frequency, and considering all of the failure
combinations.
Another evaluation should be performed regarding reactor network
economic profits. Selling the products and valuable byproducts is
considered as the income. The purchase cost of the reactors, cooler/
heater, separator, piping and instrumentation, land use, and the oper-
ating cost regarding heating and cooling loads, and raw materials are
the outcomes. Combining the incomes and the outcomes yields the
reactor network profits as the second objective function.
2.2. Basic definitions and concepts of multi-objective optimization
MOO refers to simultaneous optimization i.e., maximization and/or
minimization of two or more objective functions, which are often in
conflict with one another. This optimization problem can be specified
as follows:
Optimize f1(x),  f2 (x), …, fk (x)
Subject to  g i (x) ≤ 0 i= 1,2, …, ni
hi (x) = 0 i= 1,2, …, n e
xl < x < x u
Where the integer k≥ 2 is the number of objectives, x is a vector of
design or decision variables with lower (xl) and upper (xu) bounds, ni is
the number of inequality constraints, and ne is the number of equality
constraints (Gómez-Lorente et al., 2014).
Pareto curve
Nadir soluƟon
2
Fig. 1. Pareto frontier of a multi-objective optimization.
Compared to single objective problems, MOO problems are more
challenging to solve, because there is no unique solution. The results of
the MOO model are provided as a set of acceptable optimal solutions or
a non-dominated solution, which is called as Pareto front (Madetoja
et al., 2008). The Pareto frontier is bounded by an ideal objective vector
zideal and a nadir objective vector znadir , which describe the upper and
lower bounds for the objective function values of Pareto front, respec-
tively. All of the above-mentioned points have been shown in Fig. 1,
which illustrates the Pareto frontier for the optimization of two objec-
tion functions f1 and f2, which were minimized simultaneously.
Two main approaches to solve (MOO) problems can be dis-
tinguished. The first is called the Multi-Criteria Decision Making or
classical approach such as weighting method, constraint method, goal
programming, and the min-max approach. These methods are not able
to find multiple solutions in a single run. These methods change the
multi-objective problem into a single-objective problem with the cor-
responding weights based on their relative importance, and also suffer
from a drawback indicating that the decision maker must have a
thorough knowledge of ranking of objective functions. In contrast,
evolutionary algorithms (EAs) such as Non-dominated Sorting Genetic
Algorithm (NSGA) and elitist NSGA (NSGA-II), Pareto-Archived
Evolution strategy (PAEs), and Strength Pareto EA (SPEA) have been
verified successfully for solving multi-objective problems, because they
operate with a population of individuals and are also well suited to
search for multiple solutions simultaneously (Knowles and Corne,
1999). It is well proven that EAs can solve several conflicting objectives
and are also able to approximate the optimal solutions in a single run.
Among the EAs, One of the ways that has shown its effectiveness in
solving many multi-objective problems is the NSGA-II optimization
algorithm (Deb, 2005).
3. Case study
3.1. Process description
This study aimed at finding an economically optimum and in-
herently safer rector network for allyl chloride production. Allyl
chloride is manufactured by means of non-catalytic chlorination of
(1) propylene in the vapor phase. This process can cause all kind of risks
related to the flammable and toxic release. The following reactions
describe the process (Pahor et al., 2001).
• The principal reaction:
115
S. Eini et al. Journal of Loss Prevention in the Process Industries 51 (2018) 112–124

  18
Chlorine + Propene → Allyl Chloride + Hydrogen Sulphide
16

• Chlorination of allyl chloride to 1,3 dichloropropene:

Allyl chloride production rate (kmol/hr)

Reactor Volume = 1 m3
14 Reactor Volume = 10 m3
  Reactor Volume = 20 m3
Chlorine + Allyl Chloride → 1,3 Dichloropropene + Hydrogen Sulphide
12 Reactor Volume = 50 m3

• Chlorination of propene to 1,2 dichloropropane: 10

  8
Chlorine + Propene → 1,2 Dichloropropene
6
These reactions could be mathematically represented by the non-
isothermal Van de Vusse reaction scheme that involves four species. 4
The consecutive reaction:
2
k1 k2
A→B⎯→
⎯ C 0
0 100 200 300 400 500 600 700 800 900 1000
And the parallel reaction: Temperature (C)

k3 Fig. 3. Effect of reactor temperature on allyl chloride production rate.

2A⎯→
⎯ D
where letters A-D represent propene, allyl chloride, 1,3-di-
chloropropene and 1,2-dichloropropane, respectively, and the reaction
rates are k1,0 = 1.5 e+6 s−1, k2,0 = 4.4 e+8 s−1, k3,0 = 100 lit.
mol−1. s−1 and E1 = 66271 J/mol, E2 = 99410 J/mol, E3 = 33140 J/
mol.
The supposed superstructure of the reactor network is illustrated in
Fig. 2. Isothermal plug flow reactors (PFR) are considered in this su-
perstructure using a bypass stream. The stream that enters PFR can
either be preheated or be precooled depending on the reactor tem-
perature. Also, because of the isothermal operation in the reactors,
there are heating/cooling considerations in the reactors. The reaction
sets and the superstructure are simulated in ASPEN HYSYS process si-
mulator.
A mixture of chlorine and propene with a flow rate of 60 kmol/h
and equimolar composition is assumed as the process feed
(Temperature: 35 °C; Pressure: 200 kPa). It is supposed that all of the
unreacted chlorine and propylene is recycled and is added to the inlet
feed. The pressure drop along the process is considered to be negligible.
In order to consider Intensification strategy of ISD, feed is divided and
similar branches are supposed to process the divided feed. Three re-
actors are considered in each branch. Also, it is assumed that all of the
branches are completely similar. Consequently, the decision variables
are as follows:
✓ Bypass fraction of any reactor
✓ Temperature of any reactor
✓ Reactors volume
✓ Number of branches in parallel
It is obvious that these factors affect process profit as well as
inherent safety level of the plant. The second decision variable can be
interpreted in connection with “Moderate” strategy of inherent safety.
Both of the third and fourth variables are related to “Minimize” strategy
of inherent safety. Therefore, there are ten decision variables in the
system: (1) Number of branches in parallel, (2) Bypass fraction of the
first reactor, (3) Temperature of the first reactor, (4) Volume of the first
reactor, (5) Bypass fraction of the second reactor, (6) Temperature of
the second reactor, (7) Volume of the second reactor, (8) Bypass frac-
tion of the third reactor, (9) Temperature of the third reactor, (10)
Volume of the third reactor. In the preceding sections, these variables
were indicated using letters X1 – X10.
For the feed conditions that were mentioned previously and a PFR
with different rector volume, the effect of temperature on the allyl
chloride production rate is shown in Fig. 3. As it can be concluded from
this figure, for a reactor (with the volume lying between 1 m3 to 50 m3),
the maximum production rate of the allyl chloride is obtained in a
temperature between 200 °C to 600 °C. Therefore, in this study, the
range of the change of reactor temperature was supposed to be between
200 °C to 600 °C.
3.2. Risk objective function
In this study, risk level was considered as one of the objective
functions that should have been minimized. The primary stages of risk
assessment are hazard identification and scenario selection. In the allyl
chloride synthesis process, both toxic and flammable chemicals are
available. The toxicity hazards arise from the presence of chlorine and
hydrogen chloride. Also, release of propene (as a highly flammable
material) may lead to fire/explosion hazards. Because of the largest

Fig. 2. Intended superstructure of reactor net-

work.

116
S. Eini et al.
mass inventory in the reactors in this study, only reactors were con-
sidered as the release sources.
In the present study, several potential damage receptors were con-
sidered as follows:
• 10 people in the distance of 20 m away from the hazards source.
• 30 people in the distance of 200 m away from the hazards source.
• 350 people in the distance of 4000 m away from the hazards source.
• A building (B4-Type according to API definition (API, 1995)) at a
distance of 300 m downwind that had 20 occupants.
In this study, the individual risk (IR) was used as a measure of the
risk. Individual risk produced by an accident can be calculated using
Equation (2):
IRi = fi × Ci (2)
where:
IRi :  outcome risk of accident "i" ,
fi :  frequency of occurence of accident "i",
Ci :  consequence of accident "i".
The risk associated with reactor “r” (IRr) was determined using
Equation (3), in which all of the possible accidents are summed for
reactor “r”:
m ⎜
IRr = ∑ IRi
i= 1 (3)
In this equation “m” refers to total number of possible accidents.
3.2.1. Frequency estimation
Considering the frequency of a release as the initiating event, Event
Tree was used for calculating the frequency of different possible out-
comes. Multivariable mathematical event tree based on the theory of
accidents occurrence, the data from the past events, and the expert
judgment were all integrated in order to provide a comprehensive
model for estimation of frequency (Javidi et al., 2015). Upon a flam-
mable and toxic chemicals' catastrophic rupture, these outputs could be:
Vapor Cloud Explosion (VCE), Vapor Cloud Fire (VCF or Flash Fire),
Boiling Liquid Expanding Vapor Explosion (BLEVE) and toxic release.
Fig. 4 shows a typical Event Tree developed for a flammable and toxic
vapor release (Medina-Herrera et al., 2014).
By multiplying the initial event frequency by the probabilities of
occurrence in each branch, the frequency of the final events can be
calculated in Event Tree. For example, Equation (4) shows how the
frequency of a VCE as a final outcome of the event tree can be calcu-
lated according to the event tree that is shown in Fig. 4.
f VCE = frupture × Pdelayed ignition × (1 − Pimm.ignition) × PVCE (4)
−1
In this equation, frupture is rupture frequency in year . Also,
 Pdelayed ignition and Pimm.ignition are the probabilities of delayed ignition and
of immediate ignition, respectively. PVCE shows the probability of the
occurrence of VCE rather than VCF.
117
Journal of Loss Prevention in the Process Industries 51 (2018) 112–124
Table 4
Frequency data (LNE, 2009).
Loss of equipment event Frequency (year−1)
Catastrophic rupture 5E-05
Large breach 2.2 E −03
Medium breach 2.2 E −03
In order to specify the leakage frequency, frequency databanks such
as OGP (2010) and handbook of failure frequencies (LNE, 2009) can be
used that are quite common. Table 4 displays some generic frequency
data for pressure vessels (LNE, 2009).
3.2.1.1. Probability of immediate ignition. As it is shown in Fig. 4, after a
release, the outcome can be different depending on either the presence
or the absence of an immediate ignition source. Presence of an
immediate ignition source will lead to the formation of a BLEVE/
Fireball (Mannan, 2012).
Moosemiller (2011) proposed an equation in terms of the auto ig-
nition temperature (TAuto ignition ), the ambient temperature (T), operation
pressure (P) and the Minimum Ignition Energy (MIE) (Equation (5)).
Where TAuto ignition and T are in degrees Fahrenheit, P is in psig, and
MIE is in mJ.
⎛ −9.5 ⎛ ⎜
T ⎞⎞
⎟
⎡ P3 ⎤
1
Pimm.ignition = 1 − 5000e ⎝ TAuto ignition ⎠ + ⎢0.0024 × 2⎥
⎟
⎠ ⎣ MIE 3⎦ (5)
⎝
3.2.1.2. Probability of delayed ignition. When ignition does not occur
immediately, and there is sufficient time to form a vapor cloud, a
delayed ignition may happen. Table 5 shows the probability of delayed
ignition for various ignition sources (Uijt de Haag and Ale, 1999).
Because this study was conducted during R&D phase and conse-
quently detailed information of the plant layout was not presented, the
delayed ignition probability was conservatively considered to be equal
to 0.5.
3.2.1.3. Probability of VCE rather than VCF. Upon presence of a delayed
ignition source, two different outcomes may happen: VCE or VCF. In a
congested area or release of significant amount of flammable gas, a VCE
is more probable to happen. The probability of occurrence of VCE can
be determined in terms of the released mas using Equation (6)
(Bernechea and Viger, 2013):
PVCE = 0.024 m0.435 (6)
3.2.2. Consequence modeling
For the catastrophic release, consequence analysis involves quanti-
fication of the released mass and the effects/damages. In this study,
human life damages and structure damages were considered as im-
portant factors to be evaluated. Table 6 shows outcomes presented in
Fig. 4 and their associated effects:
For these outcomes, a detailed procedure to model the consequences
Fig. 4. Typical Event Tree developed for a flammable and toxic
vapor release (Medina-Herrera et al., 2014).
S. Eini et al. Journal of Loss Prevention in the Process Industries 51 (2018) 112–124

n
Table 5
Probability of delayed ignition in one minute for various ignition sources (Uijt de Haag IIR = ∑ Cnnf × Pnf × (IRnf )
and Ale, 1999). nf = 1 (9)

Source Probability of ignition

In Equation (9) “n ” and “nf ” represent total number of reactors in
the network and number of reactors that suffer to the similar event at
High equipment density 0.5 the same time, respectively. IRnf represents the risk produced due to the
medium equipment density 0.25 simultaneous accident of “nf” units at the same time. “Cnnf ” represents
low equipment density 0.1
confined space with- no equipment 0.02
the number of accident combinations that can occur when “nf” units
suffer an occurrence out of “n” that can be calculated using Equation
(10).
Table 6 n!
Outcomes and their physical effects. c nnf =
nf ! ( n− nf )! (10)
Outcome Physical effect

BLEVE/Fireball Thermal radiation Pnf : The probability of occurrence of this accident as independent
Blast wave
VCE Blast wave
accidents and is calculated as below:
VCF Thermal radiation 1 nf
Pnf = ⎛ ⎞
TOXIC Toxic dispersion ⎝n⎠ (11)
In the reactor network under study, different reactors have different
and the approach to estimate the related physical effects has been properties (such as volume, flammable material inventory, toxic ma-
presented by (Medina-Herrera et al., 2014; Eini et al., 2015, 2016a, terial inventory, and fluid composition). Consequently each reactor has
2016b). A Probit model is used to calculate the damage effects on its own risk level that may differ from the others. In this paper, an
human and buildings of all outcomes. Probit methods provide a gen- average risk (IRavg) is defined that can be assigned to each reactor. IRavg
eralized time dependent function, which could be used for toxic, is calculated by averaging the risks of all of the reactors in the network
thermal and blast effects (Medina-Herrera et al., 2014). (Equation (12)):
Cost of each outcome (BLEVE, VCE, VCF, or toxic) is the summation N
∑r= 1 IRr
of all of the cost damage, including human life fatality, human injuries, IR avg =
N (12)
building collapse and major building damage. The cost of the major
building damage is considered to be 70 present of the cost of the In Equation (12) N is the total number of reactor in the network and
building collapse. IRr is individual risk of r-th reactor. Finally the IR nf can be calculated
using Equation (13):
4
Ck = ∑ Costi IR nf = nf * IR avg (13)
i= 1 (7)

where 3.3. Process profit objective function

Costi : The cost of the damages (human life fatality, human injuries, In this study, process profit was another important objective func-
building collapse and building major damage) tion, which was expected to be maximized. This function is defined as
Ci : The cost of outcome “i” the total sales of products and byproducts subtracting from the expenses
such as raw material, reactor costs, utility cost and final separation cost.
Table 7 shows the cost data that have been used in the consequence This function can be shown as Equation (14).
modeling (Eini et al., 2015).
f VCE = frupture × Pdelayed ignition × (1 − Pimm.ignition) × PVCE (14)
The cost data are presented in Table 8 (Pahor et al., 2001).
3.2.3. Integrated individual risk calculation
The working time for the plant was supposed to be 8000 h/year. The
Rearranging Equation (3), the risk calculated for a reactor (as a
investment costs of a reactor and final separation were calculated using
hazard source) can be represented as below:
Equations (15) and (16), respectively (Pahor et al., 2001):
IRr = (fBLEVE*CBLEVE) + (f VCE*CVCE) + (fVCF *CVCF) + (fToxic*CToxic) (8)
Reactor cost  ⎛ $ year ⎞ = 4200 × V 0.63
The reactor network shown in Fig. 2 consists of several reactors in ⎝ ⎠ (15)
series and parallel. Hence, it is possible that a number of reactors suffer where
from the similar event at the same time. Therefore, the general equation
V: Reactor volume (m3)
that determines the total risk of network should be expanded to con-
sider all of the probability of the accident scenario combinations to
obtain the realistic results. The Integrated Individual Risk (IIR), as the Final separation cost  ⎛ $ year ⎞ = 105 × F
⎝ ⎠ (16)
overall risk level of the plant, which is one of the objective functions,
can be expressed as Equation (9) (Bernechea and Viger, 2013): Table 8
Component and utility costs (Pahor et al., 2001).
Table 7
Cost data (Eini et al., 2015). Component/utility Cost

Cost parameters Cost ($) Allyl chloride 1.717 ($/kg)

Chlorine 0.512 ($/kg)
A fatality 350,000 Propene 0.220 ($/kg)
An injury 160, 000 Cold utility 0.0025 ($/kwh)
A building 100,000 Hot utility 0.0288 ($/kwh)

118
S. Eini et al. Journal of Loss Prevention in the Process Industries 51 (2018) 112–124

Fig. 5. Flowchart of NSGA-II

where 3.4.2. Euclidian non-dimensionalization

In this method, a non-dimensionalized objective Fnij is defined as
F: Allyl chloride molar flow (mol/s) Equation (17) indicates (Sayyaadi and Mehrabipour, 2012):

It should be noted that the cost calculation was performed in 2016 Fij
Fnij = n
prices. Chemical Engineering Plant Cost Index (CEPCI) for 2001 and ∑i= 1 F2ij (17)
2016 were 394.3 and 536.4, respectively.

3.4.3. Fuzzy non-dimensionalization

3.4. Apply NSGA-II to search for the Pareto front In this method, a non-dimensionalized objective Fnij is defined as
Equation (18) shows (Sayyaadi and Mehrabipour, 2012):
In the present study, the Pareto optimal set and the corresponding
Fij − min(Fij)
Pareto frontier were achieved using NSGA-II suggested by Deb (2005). Fnij = max(Fij) − min(Fij)
(for maximization problem)
Deb developed NSGA to address three drawbacks of the methodology: max(Fij) − Fij
the lack of elitism, time complexity, and the necessity of a sharing Fnij = (for minimization problem)
max(Fij) − min(Fij) (18)
parameter for the fitness sharing system. Deb also offered NSGA-II that
comprises a second-order sorting foundation entitled crowding dis- In this work, most recognized and common types of decision making
tance, which is quicker and more adaptable than NSGA. Fig. 5 shows processes, including the Shannon entropy and fuzzy methods proposed
the algorithm evolution procedure from the first step to the end. The by Bellman-Zadeh were implemented in parallel. Moreover, the final
NSGA-II parameters for this modeling have been picked up based on optimal solution was selected based on criteria among solutions, which
logical range for each parameter, and are shown in Table 9. were suggested by these two methods. The Shannon's entropy proce-
dures use Euclidian non-dimensionalization, while Bellman–Zadeh's
fuzzy decision making method implements the fuzzy non-dimensiona-
3.4.1. Decision making in multi-objective optimization lization approach.
A procedure of decision-making is necessary in MOO to select the
final optimal solution from available solutions of Pareto front. Before 3.4.3.1. Shannon Entropy decision making. The concept of Shannon
the application of any decision making process, it is mandatory to unify Entropy that shares some intuition with Boltzmann's and some of the
the dimension and scale of the objectives (processing costs and risk mathematics developed in information theory is a useful tool to acquire
level are US $ and US $ year−1, respectively) correspondingly. Thus, weights of alternatives (Guisado et al., 2005). Considering Pij in
Euclidian and fuzzy non-dimensionalization can be used for objective decision matrix Fij with n alternatives and m objectives, the element
vectors' non-dimensionalization procedure. In those methods, the ma- of this matrix for jth objective is obtained by:
trix of objectives at various points of the Pareto frontier is denoted by Fij
where i is the index for each point on the Pareto frontier and j is the Fij
Pij = i=1, …, n j=1, …, m
∑n
i= 1 Fij (19)
index for each objective in the objectives space.

Table 9
Shannon Entropy is calculated as below:
Specified NSGA-II options for multi-objective optimization. n
1
SEj = −
ln(n)
∑ Pijln(Pij)
Specified Options Value i= 1 (20)
Population Size 30 Next step is to calculate deviation degree (dj), which shows that to
Maximum Generation 300
what extent jth objective has useful information for decision making.
Tournament Size 2
Crossover Function Intermediate Deviation degree is given by:
Crossover Probability 0.9
dj = 1 − SEj (21)
Mutation Function Constraint Dependent
Mutation Probability 0.1
Pareto Fraction 0.6 Then the weight of jth objective can be obtained based on following
equation:

119
S. Eini et al. Journal of Loss Prevention in the Process Industries 51 (2018) 112–124

dj if  Fj (x) > Fmax

wj = m
⎧1 j
∑ j= 1 dj (22) ⎪ Fj − Fmin
j
μ Fj (x) = if  Fmin
j < Fj (x) ≤ Fmax
j
⎨ Fmax
j − F min
j
A comprehensive assessment parameter is expressed as: ⎪ min
0 if  Fj (x) ≤ F j (32)
Xij = Pij × wj (23) ⎩

Then, TOPSIS method was utilized to rank the alternatives of the

new weighted normalized matrix (Xij) to find final optimal solution on
• For minimization problem
Pareto frontier. ⎧0  if Fj (x) > Fmax
j

In this method, hypothetical ideal and hypothetical non-ideal points ⎪ Fmax

j − Fj
μ Fj (x) = max − F min
if Fmin
j < Fj (x) ≤ Fmax
j
should be determined on the Pareto space. On the one hand, an ideal ⎨Fj j

point on the Pareto frontier was the point in which each objective was ⎪
1 if Fj (x) ≤ Fmin
j (33)
⎩
optimized regardless of the satisfaction of other objectives. On the other
hand, the non-ideal point was the point in the objectives' space in which
each objective had its worst value. Then, for the selection of the final • Fuzzy constraints are formulated as follows:
solution from Pareto frontier, the distance of each solution from ideal Gi (x) ≤ Gmax
i + di  ,   i = 1,2, …, k  (34)
point (di +) and non-ideal point (di −) must have been calculated sepa-
di is a subjective parameter that denotes a distance of allowable dis-
rately. Finally, the solution with minimum (di +) and maximum (di −)
placement for the bound of the jth restriction. Corresponding mem-
was chosen as finial optimum solution.
bership functions are defined using the subsequent procedure:
3.4.3.2. Bellman-Zadeh's decision making method. When applying ⎧ 0  if Gj (x) > Gmax
j
Bellman-Zadeh's approach (Arora et al., 2016), each objective ⎪ Gi (x) − Gmax
μ Gj (x) = 1 − i  if G j < Gj (x) ≤ Gmax
max
j + di
function of optimization problem was exchanged by a fuzzy objective ⎨ di
function: ⎪ 1  if Gj (x) ≤ Gmax
j (35)
⎩
Ai = {x.μ Ai} x∈ L ,  i = 1,2, …, k (24)
where μ Ai is a membership function of ith fuzzy objective function.
• Final decision is determined by the intersection of all fuzzy re-
strictions and criteria, which are accessible by means of membership
Then, fuzzy objective functions are combined to form a decision, D, functions. This problem can be reduced to the standard nonlinear
which is a fuzzy set resulted from intersection of Ai : programming problems to discover x and k values which maximize
D(x) = ∩ki= 1 Ai (x) =  mini= 1,2, … , k Ai (x) (25) k and subject as follows:

And consequently, the membership function of the intersection is
specified by:
μ D (x) = ∩ki= 1 μ Ai (x) = mini= 1,2, … , k μ Ai (x) (26)
Its use permits us to attain a solution providing the maximum de-
gree of belongingness to the fuzzy solution D:
  max D(x) = maxx∈L  mini= 1,2, … , k Ai (x) (27)
max  μ D (x) = maxx∈L  mini= 1,2, … , k μ Ai (x) (28)
And reducing the above-mentioned problem leads to a search for:
X0 = arg maxx∈L  mini= 1,2, … ,  k μ Ai (x) (29)
To obtain the above-mentioned equation output, it is necessary to
create membership function, μ Ai , i = 1, 2 … k, of objective functions, Fi
(x), i = 1, 2 … k. The membership function of the objectives, either
linear or nonlinear, can be designated depending on the context of
problem. One of possible fuzzy convolution schemes is offered below:
λ≤μ Fj , j=1,2, …, k λ≤μ Gi , i=1,2, …, n (36)
The solution of the multi-criteria problem discloses the meaning of
the optimality operator and depends on the decision-makers’ experi-
ence and understanding of the problem.
3.4.4. Integrated procedure for MOOF
Fig. 6 displays the illustrative procedure of the integrated metho-
dology, which was utilized in this study. In the first step, the super-
structure was simulated in a process simulator (ASPEN HYSYS). Risk
level and plant processing costs formulation were codified in MATLAB
as well as the MOO algorithm. The optimizer algorithm determined and
sent the decision variable to the simulator to complete a design. The
evaluation variables (such as flow, density, etc.) were extracted from
the simulator in order to calculate the corresponded value of the ob-
jective functions. At the second step, all non-dominated solutions
(Pareto frontier) were generated by using evolutionary algorithm pro-
cedure based on the NSGA-II. At the last step, final optimal points were
chosen using Shannon's entropy and Bellman-Zadeh’ fuzzy decision

• The matrix table T, which has the diagonal elements of ‘‘ideal’’ making methods. Also, in this part, the deviation index of each solution
from the ideal and non-ideal solutions is evaluated to explore the rea-
points, is defined as follows:
sonable status of various solutions obtained in dual-objective optimi-
0 0
⎡ F1 (x1) ⋯ Fn (x1) ⎤ zation and single objective optimization.
{T} = ⎢ ⋮ ⋱ ⋮ ⎥
⎢ 0 0 ⎥
⎣ F1 (x n) ⋯ Fn (x n) ⎦ (30) 3.4.4.1. Objective function, decision variables and constraints

• maximum and minimum limits for the criteria can be defined as A mathematical model of the optimization problem has three main
parts: the objective functions, the decision variables and the constraints
follows:
problem. Two objective functions were used for this study including:
Fmax
j =  maxiFi (x0i ) j= 1, …, n i= 1, …, k maximizing economic profit of the plant and minimizing risk level as a
 Fmin =  mini Fi (x0i ) j= 1, …, n i= 1, …, k measure of inherent safety level of the process. The mathematical for-
j (31)
mulations of objective functions are as follows (based on Section 3.2

• The membership functions are assumed for all fuzzy goals as fol-
and 3.3):

• Risk level:
lows:
• For maximization problem
120
S. Eini et al. Journal of Loss Prevention in the Process Industries 51 (2018) 112–124

Fig. 6. Procedure of the integrated methodology.

n
reactors were chosen based on discussions presented in Section 3.1. The
IIR = ∑ c nnf × Pnf × (IRnf )
bounds for the reactors volume and the number of the parallel branches
nf = 1 (37)
might vary from case to case. However, increasing the number of the

• Plant Profit:
parallel branches made the plant more complex.

Profit = (FlowAllyl chloride × CostAllyl chloride) − [(FlowChlorine × CostChlorine) 4. Results and discussion

+ (Flow Propene × CostPropene)] − [(Loadcooling × Costcooling )
In order to obtain an economically optimum-inherently safer reactor
+ (Loadheating × Costheating )] − [CostFinal separation ]
network, the risk level must be minimized and the process profit must
− [Cost reactors + Costheat exchangers] (38) be maximized. The optimization was carried out by a class of multi-
objective evolutionary algorithm called NSGA-II. In this regard, opti-
In this paper, three decision variables were considered for each
mization was performed with objective functions that were expressed
reactor in the superstructure.
by Equations (37) and (38) with constraints, which were expressed with
As discussed in Section 3.1, the number of the branches in parallel
Equations (39)–(42). The final optimal solutions selected by Shannon's
was considered as another decision variable. Therefore, the decision
entropy and Fuzzy decision making methods were shown on original
variables of the case study were as follows:
Pareto frontier in Fig. 7, in which the mentioned conflict between the

• T: Temperature of each reactor

considered objectives is clearly demonstrated.

• V: volume of each reactor

The ideal and nadir solutions for both objective functions

• BF: Bypass fraction of each reactor

(Processing Profit and Risk level) optimization of the reactor network
were 3.29 × 107 US $ year −1, 250.3 US $ year−1and 1.06 × 106 US
• N: Number of branches in parallel $, 929.2 US $ year−1, respectively. Therefore, the optimal solutions for
two objectives lay in 1.06 × 106 US $≤Processing Profit≤3.28 × 107
The objective functions were also solved with respect to the fol-
US $ and 250.3 US $ year−1≤ Risk level ≤929.2 US $ year−1.
lowing constraints:
It is worth mentioning that, If only the processing profit was con-
200°C ≤ T≤ 600°C (39) sidered as the objective function (single-objective optimization), the
design point B, would be chosen as the optimal design point of the
1m3 ≤ V≤ 50m3 (40)
system, while the design point A indicated the optimum system per-
0 ≤ BF ≤ 1 (41) formance considering risk level as the only objective function. As shown
in this figure, when the processing profit increases from 1.06 × 106 US
1 ≤ N≤ 20 (42)
$ to 3.24 × 107 US $, the risk level rises moderately. But infinitesimal
The lower and upper restriction values on the temperature of the increment in the processing cost from 3.24 × 107 US $ to 3.29 × 107

121
S. Eini et al. Journal of Loss Prevention in the Process Industries 51 (2018) 112–124

1000 methods are less compared with deviation indexes (0.503 and 0.497) in
single objective optimization. Therefore, the final optimal solution se-
Nadir solution B
900 lected by Shannon's entropy decision makings is most preferred, be-
cause it shows the minimum deviation index from ideal solution in
optimization procedure. The last row of Table 10 represents the nu-
800
merical values of optimum decision variables, which corresponded to
Risk Level (US $/year)

the coordinate that was determined by aforementioned decision making

700 method.
According to the results presented in Table 10, the reactor network
600 selected by the MOO algorithm that is used in this paper is shown in
Shannon Fig. 8.
solution
500 The Simplification strategy is related to the issues such as control
loops, monitoring systems, and etc., which are almost specified during
detailed design phase. The present study had focused on the pre-
400 liminary stage of the process design in which a basic design of the
A Fuzzy process was generated.
300 solution One of the best method to evaluate the simplicity of a process during
Ideal solution the basic design phase is experts' judgement. The sensitivity-analysis is
200 one of the best tools for the judgement. Therefore, in the present re-
0 0.5 1 1.5 2 2.5 3 3.5 search study, the number of branches (one of the decision variables)
Processing Profit (US $) 7
was changed as the parameter of the sensitivity analysis. Then, for a
x 10
fixed value of this parameter, other decision variables were optimized
Fig. 7. Pareto optimal frontier in the objectives space. using the multi-objective optimization algorithm. Indeed, the approach
was an integrated optimization-sensitivity analysis procedure in which
US $ leads to drastic increment in the risk level of the process. This is a decision variable (number of branches in the present study) was kept
due to the fact that by adding a parallel branch of the reactors, the risk fixed while other variables were optimized. This procedure provided a
decreases because of the lowering flow rate in each branch and con- better understanding of how the number of branches affected the value
sequently lower inventory of the reactors. However, by changing the of the objective functions. Finally, the judgment could be done con-
number of the branches, the risk level reduction is more rapid than the sidering the changes of the reactor network's size and the improve-
increase of processing costs. By increasing the number of branches, the ment/worsening of the objective functions. The results of this proce-
risk level continues to decrease, but the reduction rate decreases as the dure (for five fixed values of the number of branches) are presented in
number of paths increases. Table 11.
In order to make a distinction between different results achieved in As a result of the procedure, it can be seen that for a network with
this study, the deviation index of non-dimensional results found by two three branches in parallel (Case 3), the optimal design has only one
different decision making methods from the ideal value was calculated reactor in each branch. Therefore, the total network consists of three
as Equation (43) indicates (Kumar et al., 2015): reactors in parallel. Although the process profit of this design is lower
than the optimum design reported in Table 10 by 57 percent, this
dideal
d= network is very simple and has lower risk level by a factor of 33 per-
dnadir + dideal (43a) cent.

m 2 m 2
dideal = ∑ j= 1 (Fj − Fideal
j ) , dnadir = ∑ j= 1 (Fj − Fnadir
j )   5. Conclusion
(43b)

Deviation index of optimal solution obtained by Shannon's entropy
and Fuzzy method with single optimization approach are presented in
last column of Table 10. Shannon's entropy decision making method
yields lower value of deviation index as 0.176 with dual objective op-
timization. Moreover, deviation indices (0.176 and 0.179) obtained in
dual objective optimization with the above-mentioned decision making
In the present study, a reactor network system was modeled to
produce allyl chloride. Moreover, this system was analyzed from both
inherent safety and economic viewpoints. Employing genetic algorithm
technique, the multi-objective optimization of the system was per-
formed to obtain the optimum design parameters of the system with
regard to both inherent safety and economics. The risk posed by the

Table 10
Comparison between optimal solutions for dual-objective and single-objective optimizations.

Optimization algorithms Decision makings Objectives Deviation index

Processing profit (non-dimensional) Risk level (non-dimensional) d

NSGA − II Shannon Entropy 0.2317 0.145 0.176

Fuzzy Bellman − Zadeh 0.2223 0.1436 0.179
Single objective optimization based on maximum Processing profit 0.2343 0.3143 0.503
Single objective optimization based on minimum Risk level 0.0076 0.085 0.497
Ideal solution 0.2343 0.085 0
Nadir solution 0.0076 0.3143 1

Optimum design variables and objective functions based on Shannon entropy decision making

X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 Processing profit Risk level

16 0.04 375.4 3.7 1.00 – – 0.20 391.8 3.9 3.24 × 107 US $ 428.8 US $ year−1

122
S. Eini et al. Journal of Loss Prevention in the Process Industries 51 (2018) 112–124

Fig. 8. Optimal reactor network.

Table 11
Comparison of optimization results for five different cases.

Optimum point characterization Case 1 Case 2 Case 3 Case 4 Case 5

X1 1 2 3 4 5
X2 0.02 0.02 1.00 0.02 0.05
X3 376.9 368.8 – 335.6 355.1
X4 5.6 5.1 – 3.6 2.9
X5 0.82 0.75 0.01 0.95 0.73
X6 316.5 329.8 201.4 338.2 390.7
X7 3.6 2.8 3.4 3.3 1.7
X8 0.17 0.11 1.00 0.23 0.00
X9 366.9 350.4 – 356.7 403.3
X10 4.1 3.9 – 4.1 3.2
Processing profit (US $ year−1) 1.23 × 107 1.79 × 107 1.39 × 107 2.29 × 107 2.45 × 107
Change in the processing profit compared with the optimal case of Table 10 −62% −45% −57% −29% −24%
Risk level (US $ year−1) 707.86 660.07 285.46 622.27 614.12
Change in the risk level compared with the optimal case of Table 10 +65% +54% −33% +45% +43%

hazards in the network (considering both consequences and frequencies
of the probable accidents) and the process economic profit (including
capital and maintenance costs of system components, the operational
cost, and the profit due to selling of the product) were considered as the
objective functions in the optimization procedure. Results of the opti-
mization showed that the optimum reactor network with regard to both
inherent safety and economics had a large number of reactors in the
network that led to a highly complex system. This complexity resulted
in more process control difficulties. Also, this result was in conflict with
the simplification strategy of inherent safety design. To overcome this
problem, an integrated optimization-sensitivity analysis procedure was
performed that provided a better understanding of how the number of
branches affects the objective functions. This procedure can be con-
sidered as a tool in order to judge the process simplicity. The results
approved that the methodology can lead to an inherently safer design
considering the economic feasibility of a process.
References
API, 1995. Recommended Practice 752: Management of Hazards Associated with
Location of Process Plant Buildings, second ed. American Petroleum Institute,
Washington DC.
Arora, R., Kaushik, S.C., Kumar, R., Arora, R., 2016. Multi-objective thermo-economic
optimization of solar parabolic dish Stirling heat engine with regenerative losses
using NSGA-II and decision making. Int. J. Elec. Power 74, 25–35.
Bernechea, E.J., Viger, J.A., 2013. Design optimization of hazardous substance storage
facilities to minimize project risk. Saf. Sci. 51, 49–62.
CCPS, 2009. Inherently Safer Chemical Processes: a Life Cycle Approach, second ed. John
Wiley & Sons, Hoboken, NJ.
Deb, K., 2005. Multi-objective optimization. In: Burke, E., Kendall, G. (Eds.), Search
Methodologies. Springer, US, pp. 273–316.
Eini, S., Abdolhamidzadeh, B., Reniers, G., Rashtchian, D., 2015. Optimization procedure
to select an inherently safer design scheme. Process Saf. Environ. Protect. 93, 89–98.
Eini, S., Shahhosseini, H.R., Javidi, M., Sharifzadeh, M., Rashtchian, D., 2016a. Inherently
safe and economically optimal design using multi-objective optimization: the case of
a refrigeration cycle. Process Saf. Environ. Protect. 104 (Part A), 254–267.
Eini, S., Shahhosseini, H.R., Delgarm, N., Lee, M., Bahadori, A., 2016b. Multi-objective
optimization of a cascade refrigeration system: exergetic, economic, environmental,
and inherent safety analysis. Appl. Therm. Eng. 107, 804–817.
Gómez-Lorente, D., Triguero, I., Gil, C., Rabaza, O., 2014. Multi-objective evolutionary
algorithms for the design of grid-connected solar tracking systems. Int. J. Elec. Power
61, 371–379.
Guisado, J.L., Jiménez-Morales, F., Guerra, J.M., 2005. Application of shannon's entropy
to classify emergent behaviors in a simulation of laser dynamics. Math. Comput.
Model. 42, 847–854.
Javidi, M., Abdolhamidzadeh, B., Reniers, G., Rashtchian, D., 2015. A multivariable
model for estimation of vapor cloud explosion occurrence possibility based on a
Fuzzy logic approach for flammable materials. J. Loss Prevent. Proc. 33, 140–150.
Khan, F.I., Amyotte, P.R., 2005. I2SI: a comprehensive quantitative tool for inherent
safety and cost evaluation. J. Loss Prevent. Proc. 18, 310–326.
Khan, F., Rathnayaka, S., Ahmed, S., 2015. Methods and models in process safety and risk
management: past, present and future. Process Saf. Environ. Protect. 98, 116–147.
Kletz, T.A., Amyotte, P., 2010. Process Plants: a Handbook for Inherently Safer Design.
CRC Press.
Knowles, J., Corne, D., 1999. The Pareto archived evolution strategy: a new baseline
algorithm for multiobjective optimization. In Proceeding of the 1999 Congress on
Evolutionary Computation. IEEE Service Center 98–105.
Kumar, R., Kaushik, S.C., Kumar, R., Hans, R., 2015. Multi-objective thermodynamic
optimization of an irreversible regenerative Brayton cycle using evolutionary algo-
rithm and decision making. Ain Shams Eng. J. 7.2, 741–753.
LNE, 2009. Handbook Failure Frequencies for Drawing up a Safety Report. Brussels.
Madetoja, E., Ruotsalainen, H., Monkkonen, V.-M., Hamalainen, J., Deb, K., 2008.
Visualizing multi-dimensional Pareto-optimal fronts with a 3D virtual reality system.
In: Proceedings of the International Multi Conference on Computer Science and
Information Technology. POL, Wisla, pp. 907–913.
Mannan, M.S., 2012. Lees' Loss Prevention in the Process Industries: Hazard

123
S. Eini et al.
Identification, Assessment and Control. Butterworth-Heinemann.
Medina-Herrera, N., Jiménez-Gutiérrez, A., Mannan, M.S., 2014. Development of in-
herently safer distillation systems. J. Loss Prevent. Proc. 29, 225–239.
Moosemiller, M., 2011. Development of algorithms for predicting ignition probabilities
and explosion frequencies. J. Loss Prevent. Proc. 24, 259–265.
OGP, 2010. Report No. 434-1: Risk Assessment Data Directory: Process Release
Frequencies. International Association of Oil and Gas Procedures, Brussels.
Pahor, B., Kravanja, Z., Iršič Bedenik, N., 2001. Synthesis of reactor networks in overall
process flowsheets within the multilevel MINLP approach. Comput. Chem. Eng. 25,
765–774.
Patel, S.J., Ng, D. Ng, Mannan, M.S., 2010. Inherently safer design of solvent processes at
the conceptual stage: practical application for substitution. J. Loss Prevent. Proc. 23,
483–491.
124
Journal of Loss Prevention in the Process Industries 51 (2018) 112–124
Reniers, G.L.L., Audenaert, A., 2009. Chemical plant innovative safety investments de-
cision-support methodology. J. Saf. Res. 40, 411–419.
Sayyaadi, H., Mehrabipour, R., 2012. Efficiency enhancement of a gas turbine cycle using
an optimized tubular recuperative heat exchanger. Energy 38, 362–375.
Shariff, A.M., Rusli, R., Leong, C.T., Radhakrishnan, V.R., Buang, A., 2006. Inherent safety
tool for explosion consequences study. J. Loss Prevent. Proc. 19, 409–418.
Shariff, A.M., Zaini, D., 2010. Toxic release consequence analysis tool (TORCAT) for in-
herently safer design plant. J. Hazard Mater. 182, 394–402.
Thiruvenkataswamy, P., Eljack, F.T., Roy, N., Mannan, M.S., El-Halwagi, M.M., 2016.
Safety and techno-economic analysis of ethylene technologies. J. Loss Prevent. Proc.
39, 74–84.
Uijt de Haag, P., Ale, B., 1999. Guideline for Quantitative Risk Assessment (Purple Book).
Committee for the Prevention of Disasters, The Hague, NL.