You are on page 1of 12

Chemical Engineering Journal 431 (2022) 133244

Contents lists available at ScienceDirect

Chemical Engineering Journal


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

Deep neural network-based optimal selection and blending ratio of waste


seashells as an alternative to high-grade limestone depletion for SOX
capture and utilization
Jonghun Lim a, b, 1, Soohwan Jeong a, c, 1, Junghwan Kim a, *
a
Green Materials and Processes R&D Group, Korea Institute of Industrial Technology, 55, Jongga-ro, Jung-gu, Ulsan, 44413, Republic of Korea
b
Department of Chemical and Biomolecular Engineering, Yonsei University, 50, Yonsei-ro, Seadaemun-gu, Seoul, 03722, Republic of Korea
c
Department of Computer Science Engineering, Chungnam National University, 99, Daehak-ro, Yuseong-gu, Daejeon, 34134, Republic of Korea

A R T I C L E I N F O A B S T R A C T

Keywords: In wet flue gas desulfurization system, the resource depletion of high-grade limestone, used as conventional SOx
Wet flue gas desulfurization absorbent, is becoming serious for SOx capture and utilization. This paper proposes optimal selection and
SOx capture and utilization blending ratio of waste seashells as an alternative to high-grade limestone depletion using a deep neural network
Optimization
(DNN)-based surrogate model. Cost optimization proceeds as follows: data generation, data preprocessing,
Deep neural network
Limestone
development of DNN-based surrogate model, and derivation of cost optimal selection and blending ratio. First, a
Waste seashells process model is developed to generate the datasets, which are gypsum purity according to selection and
blending ratio of each seashell and limestone. In addition, a mathematical model is proposed to calculate the
total annualized cost (TAC) considering the pretreatment cost of seashell, and the TAC is added to the datasets to
predict the gypsum purity as well as TAC. Second, the generated datasets are preprocessed to intensify prediction
performance of the DNN-based surrogate model using the z-score normalization method. Third, a DNN-based
surrogate model is developed to predict the gypsum purity and TAC according to the selection and blending
ratio. Finally, the cost optimal selection and blending ratio are derived from 2.4 billion data generated by the
developed DNN-based surrogate model under two constraints: gypsum purity and total SOx absorbent con­
sumption. As a result, the derived selection and optimal blending ratios are low-grade limestone (80.86%), oyster
shells (10.78%), scallop shells (0.216%), cockle shells (0.323%), clam shells (2.426%), and mussel shells
(5.391%), reducing the TAC by US$788,469.

1. Introduction absorbent, whereas low-grade limestone (<94 wt% CaCO3) is not used
in view of its relatively large contents of impurities, such as MgCO3,
Sulfur dioxide (SO2) is a major air pollutant emitted by coal-fired SiO2, Al2O3, and Fe2O3. However, in South Korea, high-grade limestone
thermal power plants and a major contributor to acid rain, photo­ accounts for only ~ 20% of the total limestone reserves and is usually
chemical smog, and ozone depletion [1,2]. In most thermal power located at large depths, which makes it expensive and requires the need
plants, SOx capture and utilization are achieved through wet flue gas for alternative CaCO3 sources.
desulfurization (WFGD), during which SO2-containing flue gas is
brought into contact with an alkaline aqueous slurry containing ground
limestone [3]. This method exhibits a high desulfurization efficiency 1.1. Literature review
and produces gypsum (CaSO4), which is subsequently dehydrated and
used as a component of pigments, casts, and building materials [4,5]. The use of abundant low-grade limestone as an alternative SO2
Given that marketable gypsum should have a purity of ≥ 93 wt%, only absorbent for WFGD system has been actively studied to mitigate high-
high-grade limestone (≥94 wt% CaCO3) is currently used as an SO2 grade limestone depletion. Shobhana et al. attempted to lower the silica
content of low-grade limestone under different conditions, showing that

* Corresponding author.
E-mail address: kjh31@kitech.re.kr (J. Kim).
1
These authors contributed equally as a first author.

https://doi.org/10.1016/j.cej.2021.133244
Received 14 September 2021; Received in revised form 22 October 2021; Accepted 25 October 2021
Available online 30 October 2021
1385-8947/© 2021 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
J. Lim et al. Chemical Engineering Journal 431 (2022) 133244

Table 1 which increase expenses. Thus, the proposed solutions are difficult to
CaCO3 contents and annual production of waste seashells as well as works apply to the commercial plant because they require high energy and
probing their applications as SO2 absorbents. additional equipment installation.
Classification CaCO3 content Annual Research on seashells as SO2 - Depletion of high-grade limestone due to low utilization ratio of
[wt%] [12,13] production [t/ absorbents low-grade limestone as a substitute
y] Blending technology with high-grade limestone to solve the cost
Oyster shells 96.00 243,068 Jung et al., 2007 [14], Kim problem of the pretreatment process for the upgrade of low-grade
et al., 2019 [15], Lim et al., limestone is an alternative that allows the reduction in the consump­
2021 [4]
tion of high-grade limestone by about 14,104 ton/y. However, the high-
Scallop shells 98.00 74.67 Lu et al., 1998 [16]
Clam shells 96.80 4,803 Jung et al., 2007 [17] grade limestone consumption of the proposed blending technology is
Mussel shells 97.13 47,715 Mahidin et al., 2016 [18] still high at 10,814 ton/y, 43.4% of the total limestone consumption;
Cockle shells 95.60 324.72 Kim, 2017 [19] hence, a new alternative is necessary.
To address these challenges, using waste seashells to absorb SO2 as
high-grade limestone substitutes is a proper solution for a wet flue gas
the purity of such limestone can be increased from 36.8 to 47.3 wt% at a
desulfurization system. As waste seashells have a CaCO3 content of
superficial air velocity of 1.25 cm/s, froth depth of 25 cm, and collector
95.6–98 wt%, which is higher than that of high-grade limestone, they
dosage of 1.25 kg [6]. Kim et al. revealed that low-grade limestone can
are a resource that can be utilized as an SO2 absorbent. Recycling of
be refined by calcination (which affords CaO) and the subsequent
waste seashells may present a breakthrough for the limestone resource
exploitation of grindability differences between CaO and impurities,
depletion problem. In fact, research is underway to apply various waste
achieving a CaO content increase from 41.03 to 49.13 wt% [7]. Ahn
seashells to the desulfurization process. Table 1 shows the calcium
et al. demonstrated that calcined and ground low-grade limestone can be
carbonate content, amount of waste according to each seashell, and
subjected to magnetic separation and air classification to remove Fe2O3
research on seashells as SO2 absorbents.
and SiO2, respectively, and managed to increase CaCO3 purity from 84.1
As waste seashells are a by-product of the fishing industry, their
to 94.7 wt% [8]. Lee et al. showed that the color of low-grade sintered
annual production is strongly dependent on factors such as climate
limestone can be used to select optimal sintering, grinding, and classi­
change and buyer demand, i.e., the use of only one waste seashell type as
fication parameters, and thus, reduce the energy cost of limestone
a high-grade limestone substitute may result in feedstock availability
upgrading [9]. Tsunekawa et al. attempted to purify limestone by
problems. Therefore, to mitigate high-grade limestone depletion, one
removing impurities via reverse and carrier flotation, showing that the
can use a combination of all available types of seashells to increase their
highest removal efficiency of 65.7% was achieved for reverse flotation
recycling efficiency and availability as high-grade limestone substitutes.
with kerosene [10]. Lim et al. suggested that low-grade limestone can be
Before waste seashells can be used for SO2 absorption, they should be
mixed with high-grade limestone without pretreatment and optimized
pretreated to remove salt and organic matter. Therefore, to promote the
the corresponding mixing ratio, revealing that the fraction of low-grade
utilization of waste seashells as a substitute for high-grade limestone,
limestone in the mixture can be as high as 56.6%, which allows high-
one should optimize the limestone-to-seashell blending ratio and the
grade limestone consumption to be reduced by ~ 14,104 t/y [11].
cost of seashell pretreatment under the constraints of gypsum purity and
desulfurization efficiency.
1.2. Contribution In this study, we suggested the optimal selection and blending ratio
of each waste seashell as an alternative to high-grade limestone deple­
Despite many contributions to increase the utilization of low-grade tion for SOX capture and utilization using a deep neural network (DNN)-
limestone against high-grade limestone depletion, several challenges based surrogate model. The aim of this work is to maximize the utili­
still remain to be solved and are elaborated as follows: zation of waste seashells as SO2 absorbent which landfilled annually,
- Expensive pretreatment of low-grade limestone causing environmental problems, such as bad odor generation as well as
The improvement in the low-grade limestone quality requires sin­ to solve the resource depletion of high-grade limestone.
tering at a temperature of 900–1100 ◦ C for 2–3 h and installation of The novelty of our work can be summarized as follows.
additional process, such as a furnace, grinder, classifier, and column,

Fig. 1. Simplified diagram of WFGD process.

2
J. Lim et al. Chemical Engineering Journal 431 (2022) 133244

Fig. 2. Illustration of the employed cost optimization procedure.

• This work is the first attempt to recycle waste seashells for SOx in the wet ball mill. The mill slurry containing only finely ground
capture and utilization, and thus, mitigate the depletion of high-grade limestone enters the limestone slurry storage tank and is mixed with
limestone resources. water upon continuous stirring using an agitator to prevent the lime­
• The DNN-based cost optimization of WFGD systems is used as a stone from precipitating and clogging the pipe. At this time, CaCO3 is
novel approach to solve a mixed integer non-linear programming partly dissolved (Eq. 1) [21]:
(MINLP) problem, namely, the optimization of blending ratio for five
CaCO3 + H2 O → Ca2+ + HCO−3 + OH − . (1)
seashell types within realistic constraints. The proposed DNN-based
surrogate model allows a global optimum to be established without Subsequently, the generated alkaline limestone slurry is sprayed into
the interference of local optimum problems frequently encountered in the scrubber, and desulfurization occurs through vapor–liquid contact
complex mathematical problems. with the flue gas introduced from the bottom of the scrubber. The re­
• The results enable the complete substitution of high-grade lime­ action equations for desulfurization are as follows (Eqs. 2–4) [22].
stone with mixtures of low-grade limestone and waste seashells, and
thus, contribute to the alleviation of the environmental problems posed SO2 + H2 O ↔ H2 SO3 , (2)
by such waste and the economic problems posed by the use of costly
H2 SO3 ↔ H + + HSO−3 , (3)
high-grade limestone.

HSO−3 + 0.5O2 → H + + SO2−4 . (4)


2. Process description
As the dissolution reaction of limestone continues even in the liquid
In this section, the WFGD process is described using high-grade at the bottom of the scrubber, the slurry in the lower liquid is sprayed
limestone as an SOx absorbent for SOx capture and utilization. Fig. 1 while being continuously circulated upward through the circulation
shows a simplified diagram of the WFGD process. pump. Because it takes 12–16 h for limestone to completely dissolve, the
The SO2 absorbent (limestone) is offloaded at the factory, stored in a size of the scrubber and the residence time of solid limestone must be
tank, and subsequently transferred to a wet ball mill and ground to in­ appropriately adjusted. The limestone slurry that has absorbed SO2 is
crease the surface area, and consequently, the SO2 absorption rate. collected in a liquid phase at the bottom of the scrubber, and crystalli­
Typically, limestone is ground to a particle size of 5–20 μm to ensure zation of gypsum occurs in the liquid phase. The equations are as follows
high desulfurization efficiency [20]. The resulting mill slurry, contain­ (Eqs. 5 and 6) [23].
ing both finely and coarsely ground limestone, is transferred to the
Ca2+ + SO2−4 + 2H2 O ↔ CaSO4 ⋅2H2 O, (5)
corresponding storage tank and temporarily stored in the presence of
water. To avoid spray line clogging by coarsely ground limestone, a
Ca2+ + HSO−3 + 0.5H2 O ↔ CaSO3 ⋅0.5H2 O + H + (6)
classifier is used to ensure that only finely ground limestone is fed into
the scrubber, and the limestone discarded by the classifier is re-ground

3
J. Lim et al. Chemical Engineering Journal 431 (2022) 133244

Fig. 3. WFGD process model developed in this study.

The slurry—including gypsum, generated at the bottom of the Step 4. Derivation of cost-optimal blending parameters
scrubber—is temporarily stored in the gypsum slurry storage tank after The DNN-based surrogate model was used to determine the cost-
passing through the gypsum slurry pump. Finally, the produced gypsum optimal selection and blending ratio under the constraints of gypsum
(CaSO4⋅2H2O) is dehydrated in a centrifuge to produce the final dehy­ purity and total SOx absorbent consumption (TSAC) based on ~ 2.4
drated gypsum (CaSO4). billion data points.

3. Methodology 3.2. Process model for data generation

In this section, the methodology to derive the cost optimal selection The process model developed for the generation of gypsum purity
and blending ratio of each waste seashell and limestone is described. datasets according to selection and blending ratio was based on the work
Section 3.1 describes the procedure of cost optimization and Section 3.2 by Lim et al [11]. Considering the mixing of seashells of each type with
shows the process model for data generation. In Section 3.3, the math­ limestone of each grade, a sensitivity analysis tool was used to predict
ematical model to calculate total annualized cost (TAC) is developed. gypsum purity and desulfurization efficiency for all blending ratios, and
Finally, Section 3.4 describes the DNN-based surrogate model to derive the obtained gypsum purity datasets were used to develop the DNN-
cost optimal blending ratio. based surrogate model.

3.2.1. Thermodynamic modeling and underlying assumptions


3.1. Procedure of cost optimization
The thermodynamics of our process model was described using the
ELECNRTL method, which combines the modified NRTL model and the
This section describes the optimization procedure to derive the cost
Pitzer Debye–Hückel model to complement the interrelationship be­
of optimal selection and blending ratio of each waste seashell to solve
tween ions in the former model (activity coefficient model applicable to
the resource depletion problem of high-grade limestone in a WFGD
vapor–liquid and liquid–liquid equilibrium calculations) (Eqs. 9 and 10,
system. Fig. 2 shows the procedure of cost optimization.
respectively) [24].
To derive cost optimal selection and blending ratio of each waste
∑n
seashell with limestone, a mathematical model was proposed for data j=1 τji xj Gji
∑n ( ∑n )
xG m=1 τmi xm Gmi
generation, and the cost optimization was accomplished using the sug­ lnγi = ∑n + ∑n j ij τ ji − ∑ n , (9)
k=1 xk Gki k=1 xk Gki k=1 xk Gki
j=1
gested DNN-based surrogate model. Cost optimization was conducted
according to the following procedure. (
1 ∂Gex,local
) (
1 ∂Gex,PDH
)
Step 1. Data generation lnγi = lnγlocal
i + lnγPDH
i = m
+ m
. (10)
RT ∂ni RT ∂ni
A process model was developed for the WFGD system to generate
T,P,nj∕
=i T,P,nj∕
=i

15,836 gypsum purity datasets according the selection and blending The WFGD process was modeled under the following assumptions.
ratio. In addition, the total annualized cost (TAC) was calculated using a
mathematical model considering the seashell pretreatment cost, and • The process was in a steady state.
added to the above datasets. • The flue gas contained CO2 (72 vol%), N2 (14 vol%), O2 (5.9 vol%),
Step 2. Data preprocessing H2O (5.6 vol%), Ar (2.39 vol%), SO2 (0.07 vol%), NO2 (0.0051 vol
Given the significant scale variation in the extracted datasets, they %), and NO (0.032 vol%) [25].
were normalized using the z-score method to enhance the prediction • For oxidation with air, the O/SO2 molar ratio equaled 1.5 [4].
performance of the subsequently applied DNN-based surrogate model. • The Ca/S ratio was set to 1.05 [26].
Step 3. Development of DNN-based surrogate model • The decreases in desulfurization efficiency and gypsum purity owing
A DNN-based surrogate model with three hidden layers and 100 to the physical inhibition of impurities were ignored.
hidden nodes per layer was developed using the generated datasets to • The desulfurized clean gas was separated at the end of gypsum
predict gypsum purity and TAC based on seashell type and blending crystallization.
ratio. The Adam optimizer and the ReLU-ReLU-sigmoid activation
function were applied, and the generated datasets were divided into 3.2.2. Process modeling
training (60%), test (20%), and validation (20%) subsets for model In process modeling, Aspen Plus V10.0 was used for process simu­
development. lation, and the operating conditions of Dangjin Thermal Power Plant

4
J. Lim et al. Chemical Engineering Journal 431 (2022) 133244

Fig. 4. Optimization procedure to derive cost optimal selection and blending ratio with minimal TAC.

units 5 and 6 in Korea were referenced for the data required for 3.3.2. Total product cost (TPC)
modeling. Fig. 3 shows the process model of the WFGD system devel­ The TPC is the operating expense incurred during production and
oped to predict the gypsum purity and desulfurization efficiency. represents the ongoing cost of process operation, including the contri­
During actual WFGD, desulfurization through vapor–liquid contact butions of labor, raw material, and energy costs. The annual operating
occurs at the top of the scrubber, whereas gypsum crystallization occurs time was set to 365 d, and TPC was calculated as the sum of fixed
in the liquid phase at the bottom. Considering the different reaction charges (FC), direct production cost (DPC), plant overhead cost (OVHD),
phases for each location, we divided the scrubber process into two steps. and general expenses (GE) [32].
The clean (SO2-free) gas was designed to be separated at the rear end. In
TPC = FC + DPC + OVHD + GE (15)
addition, to demonstrate the reliability of the results derived using the
developed model, 95.5% of CaCO3 was dissolved in a limestone storage The FC is incurred regularly, regardless of business size, and can be
tank (LSST) to satisfy the actual content of unreacted limestone and calculated as the sum of local tax (Clocal tax) and insurance (Cinsurance)
purity of the gypsum derived in the experimental study [27]. Further costs [32].
details are provided in the supplementary material.
FC = Clocal tax + Cinsurance . (16)
3.3. Mathematical model used for TAC calculation The DPC directly affects production and can be calculated by sum­
ming the costs of utilities such as raw materials, water, and electricity
This section describes a mathematical model used for TAC calcula­ [32].
tion according to the selection and blending ratio of each waste seashell
DPC = Craw + Cwater + Celectricity + Cmaintenance + Clabor + Csupervision
to limestone, derived by the process model. The TAC is the annual cost, materials

composed of the annualized capital cost and annual operating and + Coperating supplies + Claboratory . (17)
maintenance cost. In general, the TAC is calculated by sum of equivalent OVHD includes the costs of factory rental, advertising, insurance,
annual cost (EAC) and total product cost (TPC), as follows [28]: and repair, i.e., the ongoing costs of process operation, and was herein
TAC = EAC + TPC (11) set to 60% of the sum of maintenance, labor, and supervision costs [32].
( )
OVHD = 0.6 × Cmaintenance + Clabor + Csupervision (18)
3.3.1. Equivalent annual cost (EAC)
The EAC is the annual cost of owning and operating an asset over its Finally, general expenses (GE) are the costs of organizing and con­
entire lifespan and is calculated by dividing the net present value (NPV) trolling general activities (as opposed to those related to a particular
of a project by the current annuity factor (AF) [29]. product or service) and are incurred as part of the day-to-day operation
[32].
NPV
EAC = (12)
AF GE = Cadministrative + Cmarketing + CR&D , (19)
Herein, the NPV was defined as the total capital investment (TCI), as
where, Cadministrative, Cmarketing, and CR&D are administrative, marketing,
it is the cost incurred during the construction and installation of the
and research and development costs, respectively.
initial process and includes the costs of equipment, land, building, offsite
facilities, supporting facilities, and installation. Generally, TCI is
calculated by summing the fixed capital cost (FCI), start-up cost (SUC), 3.4. DNN-based surrogate model used to derive cost-optimal blending
and working capital investment (WCI) [30]. ratios

TCI = FCI + SUC + WCI (13)


The DNN-based surrogate model used to predict the gypsum purity,
The AF, used to calculate the present value of a series of future an­ TAC, and hence, cost-optimal blending ratios was developed using the
nuities, is defined as datasets generated by process and mathematical models. By definition, a
[ ] surrogate model is a small extension of analytical models aimed to
1− 1 achieve close-to-real input and output characteristics when more exact
(1+RP)NP
modeling is time-consuming and costly, or when measurements are
AF = , (14)
RP associated with certain difficulties (e.g., computer simulation) [32–34].
Fig. 4 shows the simplified optimization procedure to derive optimal
where RP is the rate per period (5%), and NP is the number of periods
cost selection and blending ratio of each waste seashell and limestone
(20 y) [31].
with minimal TAC. Specifically, 15,836 values of TAC and GP according
to blending ratios were selected by considering the annual production of

5
J. Lim et al. Chemical Engineering Journal 431 (2022) 133244

each seashell waste and amount of employed limestone, and these


datasets were preprocessed and used to develop the DNN-based surro­
gate model. The cost is optimized as follows. First, based on the devel­
oped surrogate model, about 2.4 billion cases of GP and TAC according
to selection and blending ratio are predicted. Second, the constraint-
unsatisfied datasets of the predicted 2.4 billion datasets are deleted.
Finally, the optimum blending ratio (i.e., that affording the minimum
TAC) is determined under the constraints of gypsum purity and total
absorbent consumption.

3.4.1. Conventional approaches to solve the optimization problem


The optimization problem in this study is MINLP with seven decision
variables, namely high- and low-grade limestone and five types of sea­
shells. In general, mathematical programming is conventionally
employed to solve process optimization problems. However, several
limitations and challenges of the conventional mathematical program­
ming exist while solving the optimization problem, and are elaborated
Fig. 5. Structure of a typical DNN.
as follows. First, if the optimization problem to be solved is complex
with many decision variables, it is easy to fall into the local optimum.
Thus the complex nature of this problem makes it difficult to find the
global optimum among the numerous combinations of blending ratios
[33]. Second, in conventional mathematical programming, there are
limits to the calculation time owing to realistic computing power. Many
optimization problems are often complicated and based on the solution
of complex sets of differential equations, and a lot of time is spent
solving the optimization problem. Finally, there is a limit to the
complexity of constructing a mathematical model for mathematical
programming and the reproducibility of a functional structure. Thus,
some additional assumptions and correlations are required to obtain a
trackable solution [34].
To address these limitations, the approach based on the application Fig. 6. Structure of a DNN employing an activation function.
of artificial intelligence (AI), such as DNNs and genetic algorithms
(GAs), is a proper alternative solution to complex optimization problems √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
[34–36]. Therefore, in this study, to overcome the limitation of con­ 1 ∑n ( )2
ventional mathematical programming, a novel approach using a DNN- std(E) = vi − E (22)
(n − 1) i=1
based surrogate model was proposed to solve the MINLP problem,
namely, the determination of the cost-optimal seashell blending ratio
3.4.3. Development of DNN-based surrogate model
without any complex construction of mathematical models related to the
A DNN is based on an artificial neural network (ANN), a machine
TAC and GP, and the selection and blending ratio of each waste seashell.
learning algorithm created by a computer that mimics the principles and
In addition, the DNN-based surrogate model determines the global op­
structure of a human neural network [38,39].
timum without encountering local optimum problems frequently
A typical ANN features an input layer, hidden layer, and output
observed for complex mathematical problems very quickly.
layer, with ANNs comprising three or more hidden layers denoted as
DNNs (Fig. 5). In general, linear estimation functions of the type y = wTx
3.4.2. Data preprocessing
+ b are considered to solve linear problems in regression analysis or
The data generated in Step 1 featured a large-scale variation in
classification problems. To solve nonlinear problems, a DNN applies an
attribute values, e.g., the input of limestone and seashells ranged from
activation function to a linear estimation function.
0 to 3,710 kg/h, gypsum purity ranged from 0 to 1, and TAC exceeded
Fig. 6 shows the calculation process that determines the value of the
10,000,000. As this variation could degrade prediction performance in
unit in the next layer based on the units existing in a specific layer (input
the process of learning regression analysis using a neural network, it was
or hidden layer) in the neural network. Here, x1 to xn denote input
necessary to perform scaling, i.e., the normalization of each attribute
vectors 1–n, w1 to wn denote the weights of these input vectors, and b
[37]. Herein, we used the outlier (abnormal data point)-insensitive z-
denotes the bias added after the weighted summation of the input vec­
score normalization following a normal distribution with a mean of zero
tors. Up to this point, the results are the same as those provided by the
and a standard deviation of unity, using the mean and standard devia­
linear estimation function y = wTx + b but are then processed using the
tion of the same attribute values. The z-score was calculated as
nonlinear activation function to afford values to be used as the output or
vi − E the input of the next hidden layer. In general, a DNN uses a back-
v’i = (20) propagation process (output layer → hidden layer → input layer). The
std(E)
differential value of the activation function is used to update the weight,
where E denotes the value of a certain attribute, vi denotes the ith value and thus, reduce the error between the output of the final output layer
of E, vi ’ denotes the z-score-normalized value of vi , and E denotes the and the actual output. Therefore, for back-propagation to proceed, the
mean of E, and is calculated as activation function must be differentiable, i.e., it should not be a linear
expression or a constant. The commonly used activation functions
1∑n
E= vi (21) include ReLU, sigmoid, and tanh (Eq. 22).
n i=1
Further, std(E) denotes the standard deviation of the attribute value ReLU(x) = max(0, x); sigmoid(x) = (1 + e− x )− 1 , tanh(x)
/
and is calculated as = (ex − e− x ) (ex + e− x ). (23)

6
J. Lim et al. Chemical Engineering Journal 431 (2022) 133244

Table 2
Datasets generated using the process model and the TAC calculation model.
Case THC [kg/h] TLC [kg/h] TOC [kg/h] TSC [kg/h] TCKC [kg/h] TCC [kg/h] TMC [kg/h] GP [%] TAC [$]

1 0.000 0.000 0.000 0.000 0.000 0.000 1,237 95.67 11,255,380


2 0.000 0.000 0.000 0.000 0.000 182.8 1,237 95.72 11,330,092
3 0.000 0.000 0.000 0.000 12.36 0.000 2,473 95.68 11,737,784
⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮
15,834 3,710 3,710 3,710 8.524 24.71 182.8 1,236 83.14 15,745,989
15,835 3,710 3,710 3,710 8.524 24.71 365.5 1,236 82.99 15,802,324
15,836 3,710 3,710 3,710 8.524 37.07 548.3 3,710 80.79 16,565,118

To develop the DNN-based surrogate model, the Python version 3.7.7 performance metrics can be obtained from thousands of different values
environment and keras version 2.3.1 framework—the open-source depending on data scale, which makes it difficult to immediately
neural network library written in Python—were adopted. The datasets determine performance based on absolute values, R2 has the advantage
were randomly shuffled and split after random shuffling to ensure that of intuitively representing the performance of the regression model as a
data in multiple sections could be learned evenly for robust DNN-based relative numerical value. The performance of the regression analysis-
surrogate model development. In addition, the datasets were divided based surrogate model developed herein was evaluated using RMSE
into training (60%), test (20%), and validation (20%) sets [40,41]. and R2.
Specifically, the regression model was trained using 9,501 data entries,
and performance during training was continuously monitored using 4. Results and discussion
3,167 validation data entries. When the training was completed, the
performance of the final model was measured using 3,168 test data 4.1. Dataset generation
entries. The constructed DNN model had three hidden layers (h_1, h_2,
h_3) with 100 units in each, and it was heuristically set to develop a The datasets to be fed to the DNN for surrogate model development
high-performance surrogate model. The use of the sigmoid activation were generated using the process model and the mathematical model
function for the front hidden layer, except for the last hidden layer, was used for TAC calculation. For data generation, the usage of high-grade
likely to cause gradient vanishing during back-propagation. Accord­ limestone, low-grade limestone, oyster shells, and mussel shells was
ingly, h_1 and h_2 were set to ReLU, and h_3 was set to sigmoid. The set to 0–3,710 kg/h based on the maximum available amount of lime­
optimizer function was set to Adam because the learning rate can be self- stone and the annual amount of generated waste seashells; the usage of
adapted during the learning phase. This, therefore, reduces the effort of scallop, cockle, and clam shells was set to 0–9, 0–38, and 0–548 kg/h,
selecting an appropriate learning rate, and it is easier to tune hyper- respectively. The gypsum purity and desulfurization efficiency for the
parameters than for other optimizers [42–44]. The MSE is the 15,836 blending ratios within the above ranges were obtained using a
preferred loss function in the maximum likelihood inference framework sensitivity analysis tool. To train the DNN-based model, the TAC was
when the distribution of the target variable is normal. Thus a mean added through the corresponding mathematical model to the 15,836
squared error (MSE) suitable for regression analysis was applied as the gypsum purity datasets according to the selection and blending ratio
loss function because we scaled the dataset to follow the normal dis­ extracted by the process model. Table 2 lists the datasets derived using
tribution with z-score [45]. process and mathematical models.
In all 15,836 cases, the desulfurization efficiency equaled 99.99%,
3.4.4. Evaluation criteria for model performance and thus, satisfied the constraints. This behavior was expected as
The MSE (Eq. 23), root mean squared error (RMSE, Eq. 24), and equilibrium-based modeling was carried out assuming the proposed
coefficient of determination (R2, Eq. 25) are often used as metrics to process model to be in a steady state. In the actual process, the main
evaluate the performance of regression analysis models. For y as the parameter determining desulfurization efficiency is the Ca/S ratio,
actual value, ̂y as the value predicted by regression analysis, and n data which is the ratio between the SOx absorbent and SOx in the flue gas, and
entries, the above metrics are calculated as follows[39]. is generally set to 1.03–1.05 in WFGD systems. Considering this, the
( )2 target process was operated by setting the total consumption of the SOx
∑n
yi
i=1 yi − ̂
absorbent to 3,710 kg/h. In Entry 1 of Table 2, the total SOx absorbent
MSE = , (24) consumption is as small as 1,237 kg/h, as it only includes the total
n mussel shell consumption (TMC). As the actual desulfurization effi­
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑n ciency is smaller than 99.99%, there may be cases where the emission
RMSE = i=1 (yi − ̂ y )2
, (25) standard is not satisfied. Therefore, in this study, the SOx absorbent
n consumption was set to 3,710 kg/h as a constraint for deriving the cost-
( )2 optimal blending ratio in the DNN-based model while only considering
∑n cases satisfying the DE emission standard.
i=1 yi − ̂y i
SSR SSE
R2 = = 1− = 1− ∑n . (26)
SST SST i=1 (yi − y)2 4.2. Surrogate model performance
If RMSE is used, it is easier to interpret the error metric by converting
it back to a unit similar to the actual value than using MSE. As RMSE This section discusses the performance of the surrogate model
divides it by the square root of n, it has the advantage of being robust by developed based on 15,836 datasets for gypsum purity and TAC ac­
assigning a penalty when a large error occurs [46]. cording to each blending ratio to determine the cost optimal selection
The total sum of squares (SST) is expressed as Y − Y and denotes the and blending ratio. While training the regression analysis model with
total deviation, whereas the regression sum of squares (SSR) is expressed 9,501 training data, the model performance during training was
continuously measured with 3,167 validation data, and training was
as Y
̂ − Y and denotes the deviation explained by the regression equation.
induced in the direction of improving the performance. When training
Finally, the error sum of squares (SSE) is expressed as Y − Y
̂ and denotes
was finished, the performance of the final model was measured with
the deviation not explained by the regression equation. While other 3,168 test data.

7
J. Lim et al. Chemical Engineering Journal 431 (2022) 133244

Table 3 Case 2 is a case where only waste scallop shells are considered as a
Performance of DNN-based surrogate model [5]. substitute for high-grade limestone, and Fig. 9 shows the TAC according
Evaluation criteria Gypsum purity TAC to blending ratio of scallop shells mixed with limestone.
Given that 74.67 tons of waste scallop shells are produced annually,
MAE 0.0005 14,218
MSE 0.000001 303,804,900
RMSE 0.0010 17,430
R2 0.9990 0.9998

Table 3 summarizes the results of model performance quantification


based on the errors of the 3,168 test data entries (actual) and the data
predicted by the DNN regression model. For the 15,836 real data entries,
gypsum purity equaled 0.808–0.963 (RMSE = 0.0010), whereas TAC
ranged from 10,618,569 to 16,566,148 (RMSE = 17,429). A scale-
independent evaluation method afforded R2 values of 0.9990 and
0.9998 for gypsum purity and TAC, respectively.
Fig. 7 shows the distribution of the 3,168 actual values of gypsum
purity and TAC and the corresponding values predicted by the DNN
model, revealing that the latter model had good predictive power and
was highly reliable, i.e., it was suitable for use as a surrogate model.

4.3. Case study


Fig. 8. TAC according to blending ratio of oyster shell with limestone.
To prove the economic feasibility of the cost optimalselection and
blending ratio of each seashell and limestone proposed, five cases were
selected where only one type of seashell was mixed with limestone. The
results of this case study are compared to the derived cost optimization
results in Section 4.4.
Case 1: Blending of waste oyster shells
First, in case 1, only waste oyster shells are considered as a substitute
for high-grade limestone, and Fig. 8 shows the TAC according to the
blending ratio of oyster shells mixed with limestone.
The annual amount of waste for oyster shells is 243,068 t/y, and the
maximum mixable amount with limestone is 27,747 kg/h. Therefore,
limestone can be completely replaced. Fig. 7 shows the TAC as a function
of blending ratio of oyster shell, revealing that with increasing blending
ratio, TAC first increases owing to the large influence of pretreatment
costs at low blending ratios, decreases owing to the reduction in the raw
material cost following the replacement of limestone, and increases
again (at blending ratios above 16%) owing to the increase in pre­
treatment cost. The optimal blending ratio of obtained cost is 16%, and
the TAC here is US$11,904,367.
Fig. 9. TAC according to blending ratio of scallop shells with limestone.
Case 2: Blending of scallop shells

Fig. 7. Plot of predicted vs. actual values for gypsum purity and TAC. The red solid line corresponds to y = x, with blue dots indicating individual data points. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

8
J. Lim et al. Chemical Engineering Journal 431 (2022) 133244

Fig. 10. TAC according to the blending ratio of clam shells with limestone. Fig. 12. TAC according to the blending ratio of cockle shells with limestone.

Table 4
Results of WFGD system optimization.
Variables Description Conventional Optimal value
value (blending ratio)

THC [kg/ Total high-grade 3,710 0.000 (0.000%)


h] limestone consumption
TLC [kg/ Total low-grade 0.000 3,000 (80.86%)
h] limestone consumption
TOC [kg/ Total oyster shell 0.000 400.0 (10.78%)
h] consumption
TSC [kg/ Total scallop shell 0.000 8.000 (0.216%)
h] consumption
TCKC [kg/ Total cockle shell 0.000 12.00 (0.323%)
h] consumption
TCC [kg/ Total clam shell 0.000 90.00 (2.426%)
h] consumption
TMC [kg/ Total mussel shell 0.000 200.0 (5.391%)
h] consumption
TSAC [kg/ Total SOx absorbent 3,710 3,710
h] consumption
Fig. 11. TAC according to the blending ratio of mussel shells with limestone.
GP [%] Gypsum purity 93.90 93.00
TAC [$] Total annualized cost 12,665,991 11,877,522
the maximal consumption rate of these shells equals 8.52 kg/h, which
corresponds to a maximal blending ratio of 0.23%. In Fig. 8, it is seen
that the TAC steadily increases as the scallop shell blending ratio in­ $11,919,415.
creases to 0.23%. As the maximum mixable amount is very small, the Case 5: Blending of cockle shells
effect of the increase in the related pretreatment costs exceeds the effect Finally, case 5 is a case where only waste cockle shells are considered
of reduced limestone consumption, i.e., the mixing of scallop shells with as a substitute for high-grade limestone, and Fig. 12 shows the TAC
limestone is not economically efficient. according to the blending ratio of cockle shells mixed with limestone.
Case 3: Blending of clam shells Given that 324.7 tons of waste mussel shells are produced annually,
In case 3, only waste clam shells are considered as a substitute for the maximal blending ratio of these shells equals 1%. Therefore, like
high-grade limestone, and Fig. 10 shows the TAC according to the scallop shells, cockle shells can be blended at only small ratios, but such
blending ratio of clam shells mixed with limestone. blending is not economically efficient.
Given that 4,803 tons of waste clam shells are produced annually, the
maximal consumption rate of these shells equals 548 kg/h, which cor­ 4.4. Optimization results
responds to a maximal blending ratio of 14.78%. Fig. 8 shows that with
increasing blending ratio, the TAC first increases owing to the significant To derive optimal selection and blending ratio of waste seashells as
effect of pretreatment costs at low blending ratios, and then decreases an alternative to high-grade limestone depletion using the DNN-based
owing to the dominant effect of reduced limestone consumption. The surrogate model, the objective function was set to minimize TAC,
optimal blending ratio is determined as 14.78%, corresponding to a TAC which features the contributions of annualized capital costs and annual
of US$11,941,866. operating and maintenance costs. The equation of objective function is
Case 4: Blending of mussel shells as follows:
In case 4, only waste mussel shells are considered as a substitute for Minimize TAC
high-grade limestone. Fig. 11 shows the TAC according to the blending TAC = EAC + TPC. (27)
ratio of mussel shells mixed with limestone.
Given that 47,715 tons of waste mussel shells are produced annually, As constraints, first, the total SOx absorbent consumption (TSAC) was
the maximal consumption rate of these shells equals 5,447 kg/h, which operated at 3,710 kg/h considering the Ca/S ratio required for SO2
corresponds to a maximal blending ratio of 18.30%. The optimal absorption in the target process. Therefore, the TSAC was 3,710 kg/h,
blending ratio of obtained cost is 18.30%, and the current TAC is which was set as a constraint and composed of total seashell consump­
tion (TSC) and total limestone consumption (TLC). For TSC, the

9
J. Lim et al. Chemical Engineering Journal 431 (2022) 133244

Table 5 limestone. The TACs of each case were compared to that derived in the
Recycling efficiencies of waste seashells in Republic of Korea. cost optimal selection and blending ratio. Fig. 13 shows the decrease in
Classification Oyster Scallop Clam Mussel Cockle TAC (compared to the conventional process) for various SO2 absorbents.
shells shells shells shells shells In this comparison, cases 2 (only scallop shells) and 5 (only cockle
Amount of waste [ton/ 243,068 74.67 4,803 47,715 324.72 shells) were excluded as they exhibited an increase in TAC. As a result of
y] the comparison, the TAC became economical because it was reduced in
Consumption as high- 3,504 70.08 788.4 1,752 105.1 all cases compared to the conventional process where only high-grade
grade limestone limestone was used. The blending technology of low- and high-grade
substitutes [ton/y]
Recycling efficiency 1.442 93.85 16.42 3.672 32.37
limestone proposed in the previous study showed significantly less
[%] TAC reduction at $588,609 compared to the cases where the waste
seashells were used as a high-grade limestone substitute. On the con­
trary, when the waste seashells were used as a high-grade limestone
maximum consumption is set by referring to the annual amount of waste substitute, the TAC was reduced more, which suggested that waste
in Table 1. The equation of TSAC is as follows: seashells could be used as an economical substitute for high-grade
TSAC = TSC + TLC = 3, 710. (28) limestone. However, the annual production of each sea shell can
change critically owing to influences such as climate change or buyer
In addition, the gypsum purity (GP) must be 93 wt% or more to demand, resulting in feedstock availability problems. To solve this
achieve a quality standard. Thus, this was reflected as the constraint problem, we attempted to mix and use all five types of seashells to in­
whose equation is as follows: crease the resource recycling rate of discarded seashells, thereby
GP ≥ 93. (29) increasing their availability as high-grade limestone substitutes. The
TAC comparison results suggested that reduction in TAC under optimal
First, Table 4 presents the optimization results of the WFGD system cost selection and blending ratio obtained by mixing all five types of
via waste seashells as high-grade limestone substitutes. As a result of seashells was $788,469, showing the highest economic feasibility. This
optimization, it was possible to completely replace high-grade limestone indicated that mixing and using several different types of seashells could
by mixing 3,000 kg/h of limestone and 710 kg/h of waste seashells, with address the feedstock availability problem while demonstrating the
the TAC reduced by about $788,469. The optimal selection and blending highest economic feasibility.
ratio derived at this time was 80.86% for low-grade limestone, 10.78%
for oyster shells, 0.216% for scallop shells, 0.323% for cockle shells, 5. Conclusion
2.426% for clam shells, and 5.391% for mussel shells. The gypsum pu­
rity was 93.00%, and the total SOx absorbent was 3,710 kg/h, which In this study, we suggested the optimal selection and blending ratio
satisfied all the constraints. of each waste sea shell as high-grade limestone substitutes through cost
Second, areas of waste seashell production and landfilling are ex­ optimization of a WFGD system using a DNN-based surrogate model.
pected to experience serious environmental pollution. In Korea, This paper has two major contributions to the literature. First, this work
~295,985 tons of waste seashells are landfilled annually, causing is the first attempt to utilize each waste sea shell as high-grade limestone
environmental problems such as bad odor and water pollution. There­ substitutes simultaneously to solve the resource depletion of high-grade
fore, the recycling of waste seashells as SOx absorbents allows one to limestone in a WFGD system for SOx capture and utilization. Second, a
mitigate both the depletion of high-grade limestone resources and novel approach using DNN-based cost optimization to solve MINLP
environmental pollution. Table 5 shows the recycling efficiencies of problem is suggested, namely optimal selection and blending ratio of sea
waste seashells according to optimization results, revealing that shells within realistic constraints. Through the DNN-based surrogate
1.442–93.85% of annually produced seashell waste can be recycled. model, a global optimum is derived without local optimum problems
Thus, the use of seashells as a substitute for high-grade limestone is that are frequent for complex mathematical problems.
expected to offer great environmental benefits. The derived optimal selection and blending ratios were low-grade
Finally, to prove the economic feasibility of the cost optimal selec­ limestone (80.86%), oyster shells (10.78%), scallop shells (0.216%),
tionand blending ratio of each seashell and limestone proposed, five cockle shells (0.323%), clam shells (2.426%) and mussel shells
cases were selected where only one type of seashell was mixed with (5.391%). Thus, it is possible to completely substitute the high-grade

Fig. 13. Effect of SO2 absorbents on TAC decrease relative to conventional process.

10
J. Lim et al. Chemical Engineering Journal 431 (2022) 133244

limestone through blending with each waste sea shell and low-grade replacement, IOP Conf. Ser, Mater. Sci. Eng. 271 (2017) 012059, https://doi.org/
10.1088/1757-899X/271/1/012059.
limestone. In addition, the TAC of the optimal blending ratio could be
[13] P. Lertwattanaruk, N. Makul, C. Siripattarapravat, Utilization of ground waste
reduced by about US$788,469. seashells in cement mortars for masonry and plastering, J. Environ. Manage. 111
This study provides valuable insights for utilization of the waste sea (2012) 133–141, https://doi.org/10.1016/j.jenvman.2012.06.032.
shell as high-grade limestone substitutes in the WFGD system, making [14] J.H. Jung, K.S. Yoo, H.G. Kim, H.K. Lee, B.H. Shon, Reuse of waste oyster shells as a
SO2/NOx removal absorbent, J. Ind. Eng. Chem. 13 (2007) 512–517.
the process cost effective and environmentally safe. The cost optimal [15] S.H. Kim, B.U. Hong, J.W. Lee, W.S. Cha, K. Kim, B.K. Moon, Evaluation of SO2
selection and blending ratio of waste sea shell with limestone derived Absorption Efficiency for Calcined Oyster Shell Slurry Using a Simulated Spray
through optimization can lead to excellent effects in terms of economic Type-flue Gas Desulfurization (FGD) System: A comparative study with limestone
slurry, Econ. Environ. Geol. 52 (2019) 119–128, https://doi.org/10.9719/
and environmental improvements. Thus, we believe that our study EEG.2019.52.2.119.
makes a significant contribution to literature because it paves the way to [16] G. Lu, H. Kim, J. Yuan, I. Naruse, K. Ohtake, M. Kamide, Experimental study on
the establishment of cheaper and more eco-friendly WFGD techniques self-desulfurization characteristics of biobriquette in combustion, Energy and
Fuels. 12 (4) (1998) 689–696, https://doi.org/10.1021/ef970194c.
and presents a novel approach to solving a mixed integer non-linear [17] J.-H. Jung, Y.-S. Lee, K.-S. Yoo, H.-K. Lee, K.-J. Oh, B.-H. Shon, Reactivity of bio-
programming problem. Further, this paper will be of interest to the sorbent prepared by waste shells of shellfish in acid gas cleaning reaction, Korean
broader audience because it addresses some of its key focus areas, J. Chem. Eng. 22 (4) (2005) 566–568, https://doi.org/10.1007/BF02706644.
[18] M. Mahidin, A. Gani, M.R. Hani, M. Syukur, H. Hamdani, K. Khairil, S. Rizal,
namely, resource recovery with SOx capture and utilization. A. Hadi, T.M.I. Mahlia, Use of green mussel shell as a desulfurizer in the blending
of low rank coal-biomass briquette combustion, Makara J. Technol. 20 (2) (2016)
97, https://doi.org/10.7454/mst.v20i210.7454/mst.v20i2.3334.
Declaration of Competing Interest
[19] Y.S. Kim, Manufacturing high quality desulfrutization lime method using shell for
wet desulfurization 0093485 (2017) 1–5.
The authors declare that they have no known competing financial [20] P. Córdoba, Status of Flue Gas Desulphurisation (FGD) systems from coal-fired
interests or personal relationships that could have appeared to influence power plants: Overview of the physic-chemical control processes of wet limestone
FGDs, Fuel. 144 (2015) 274–286, https://doi.org/10.1016/j.fuel.2014.12.065.
the work reported in this paper. [21] C. Carletti, C. De Blasio, M. Miceli, R. Pirone, T. Westerlund, Ultrasonic enhanced
limestone dissolution: Experimental and mathematical modeling, Chem. Eng.
Acknowledgments Process. - Process Intensif. 118 (2017) 26–36, https://doi.org/10.1016/j.
cep.2017.04.012.
[22] B. Dou, W. Pan, Q. Jin, W. Wang, Y.u. Li, Prediction of SO2 removal efficiency for
This work was supported by the Korean Institute of Industrial wet Flue Gas Desulfurization, Energy Convers. Manag. 50 (10) (2009) 2547–2553,
Technology within the framework of the following projects: “Develop­ https://doi.org/10.1016/j.enconman.2009.06.012.
[23] K. Song, Y.N. Jang, W. Kim, M.G. Lee, D. Shin, J.H. Bang, C.W. Jeon, S.C. Chae,
ment of Global Optimization System for Energy Process [grant number Precipitation of calcium carbonate during direct aqueous carbonation of flue gas
EM-21-0022, IR-21-0029, IZ-21-0052]”, “Development of AI Platform desulfurization gypsum, Chem. Eng. J. 213 (2012) 251–258, https://doi.org/
Technology for Smart Chemical Process [grant number JH-21-0005]” 10.1016/j.cej.2012.10.010.
[24] J. Lim, J. Lee, H. Cho, J. Kim, Model development of amine regeneration process
and “Development of Energy Process Optimization Platform for Carbon with electrodialysis reclamation unit, Comput. Aided Chem. Eng. 50 (2021),
Neutrality [grant number IR-21-0033]”. https://doi.org/10.3390/min7110207.
[25] N. Meunier, S. Laribi, L. Dubois, D. Thomas, G. De Weireld, CO2 capture in cement
production and re-use: First step for the optimization of the overall process, Energy
Appendix A. Supplementary data Procedia. 63 (2014) 6492–6503, https://doi.org/10.1016/j.egypro.2014.11.685.
[26] M. Fan, R.C. Brown, Reduction of S02 in Flue Gas and Applications of Fly Ash : A
Supplementary data to this article can be found online at https://doi. Review Reduction of S02 in Flue Gas and Applications of Fly Ash, A Review (2008).
[27] A. Mizan, D. Muraskin, R. Lerose, High performance wfgd system for florina power
org/10.1016/j.cej.2021.133244. station : features operating experiences and performance results, Third Int, Symp,
2004.
References [28] J. Lim, J. Kim, Optimization of a Wet Flue Gas Desulfurization System Considering
Low-grade Limestone and Waste Oyster Shell, J. Korea Soc. Waste Manag. 37 (04)
(2020) 263–274, https://doi.org/10.9786/kswm.2020.37.4.263.
[1] Z. Li, Y. Liu, H. Wang, C.J. Tsai, X. Yang, Y. Xing, C. Zhang, P. Xiao, P.A. Webley,
[29] Y. Choi, N. An, I.l. Moon, J. Kim, Energy Optimization via Process Modification to
A numerical modelling study of SO2 adsorption on activated carbons with new rate
Maximize Economic Feasibility of the Butane Gas-Splitting Process, Ind. Eng.
equations, Chem. Eng. J. 353 (2018) 858–866, https://doi.org/10.1016/j.
Chem. Res. 59 (40) (2020) 18019–18027, https://doi.org/10.1021/acs.
cej.2018.07.119.
iecr.0c0243010.1021/acs.iecr.0c02430.s001.
[2] B.C. Bai, H.U. Lee, C.W. Lee, Y.S. Lee, J.S. Im, N2 plasma treatment on activated
[30] A. Zuorro, K. Moreno-Sader, A. González-Delgado, Economic Evaluation and
carbon fibers for toxic gas removal: Mechanism study by electrochemical
Techno-Economic Sensitivity Analysis of a Mass Integrated Shrimp, Polymers
investigation, Chem. Eng. J. 306 (2016) 260–268, https://doi.org/10.1016/j.
(Basel). 12 (2020) 2397.
cej.2016.07.046.
[31] C. Rathgeber, E. Lävemann, A. Hauer, Economic top-down evaluation of the costs
[3] R. Huang, R. Yu, H. Wu, D. Pan, Y. Zhang, L. Yang, Investigation on the removal of
of energy storages-A simple economic truth in two equations, J. Energy Storage. 2
SO3 in ammonia-based WFGD system, Chem. Eng. J. 289 (2016) 537–543, https://
(2015) 43–46, https://doi.org/10.1016/j.est.2015.06.001.
doi.org/10.1016/j.cej.2016.01.020.
[32] Y.S. Jeong, J. Jung, U. Lee, C. Yang, C. Han, Techno-economic analysis of
[4] J. Lim, H. Cho, J. Kim, Optimization of wet flue gas desulfurization system using
mechanical vapor recompression for process integration of post-combustion CO2
recycled waste oyster shell as high-grade limestone substitutes, J. Clean. Prod. 318
capture with downstream compression, Chem. Eng. Res. Des. 104 (2015) 247–255,
(2021) 128492, https://doi.org/10.1016/j.jclepro.2021.128492.
https://doi.org/10.1016/j.cherd.2015.08.016.
[5] J. Lim, Y. Choi, G. Kim, H. Song, J. Kim, Modeling of wet flue gas desulfurization
[33] Y. Ahn, J. Kim, J.-I. Kwon, Optimal design of supply chain network with carbon
process for utilization of low-grade limestone, Korean Chem. Eng. Res. 57 (2019)
dioxide injection for enhanced shale gas recovery, Appl. Energy. 274 (2020)
743–748, https://doi.org/10.9713/kcer.2019.57.5.743.
115334, https://doi.org/10.1016/j.apenergy.2020.115334.
[6] S. Dey, L. Sahu, B. Chaurasia, B. Nayak, Prospects of utilization of waste dumped
[34] J. Krzywanski, H. Fan, Y. Feng, A.R. Shaikh, M. Fang, Q. Wang, Genetic algorithms
low-grade limestone for iron making: A case study, Int. J. Min. Sci. Technol. 30 (3)
and neural networks in optimization of sorbent enhanced H2 production in FB and
(2020) 367–372, https://doi.org/10.1016/j.ijmst.2020.03.011.
CFB gasifiers, Energy Convers. Manag. 171 (2018) 1651–1661, https://doi.org/
[7] H. Kim, Method for improving the quality of low-grade limestone and the
10.1016/j.enconman.2018.06.098.
limestone, Korea Paten. 1161755 (2012) 1–21.
[35] J. Krzywanski, A general approach in optimization of heat exchangers by bio-
[8] J. Ahn, W. Jeong, Y. Yang, Manufacturing methods of high-quality lime for
inspired artificial intelligence methods, Energies. 12 (23) (2019) 4441, https://doi.
desulfurization from low-grade limestone, Korea Pat. (2015).
org/10.3390/en12234441.
[9] H. Lee, K. Kim, J. Kim, K. You, H. Lee, Breakage characteristics of heat-treated
[36] J. Krzywanski, K. Sztekler, M. Bugaj, W. Kalawa, K. Grabowska, P.R. Chaja,
limestone determined via kinetic modeling, Minerals. 8 (2018) 1–10, https://doi.
M. Sosnowski, W. Nowak, L. Mika, S. Bykuc, Adsorption chiller in a combined
org/10.3390/min8010018.
heating and cooling system: Simulation and optimization by neural networks, Bull.
[10] M. Tsunekawa, Y. Honma, K. Yoo, N. Hiroyoshi, M. Ito, Removal of trace impurity
Polish Acad. Sci. Tech. Sci. 69 (2021) 1–11. https://doi.org/10.24425/bpasts.20
from limestone using flotation techniques, Mater. Trans. 50 (1) (2009) 171–176,
21.137054.
https://doi.org/10.2320/matertrans.M-MRA2008839.
[37] S.G.K. Patro, K.K. sahu, Normalization: A preprocessing stage, Iarjset. (2015)
[11] J. Lim, Y. Choi, G. Kim, J. Kim, Modeling of the wet flue gas desulfurization system
20–22. https://doi.org/10.17148/iarjset.2015.2305.
to utilize low-grade limestone, Korean J. Chem. Eng. 37 (12) (2020) 2085–2093,
[38] J. Schmidhuber, Deep Learning in neural networks: An overview, Neural Networks.
https://doi.org/10.1007/s11814-020-0639-6.
61 (2015) 85–117, https://doi.org/10.1016/j.neunet.2014.09.003.
[12] W.A.S.B. Wan Mohammad, N.H. Othman, M.H. Wan Ibrahim, M.A. Rahim,
S. Shahidan, R.A. Rahman, A review on seashells ash as partial cement

11
J. Lim et al. Chemical Engineering Journal 431 (2022) 133244

[39] H. Kwon, K.C. Oh, Y. Choi, Y.G. Chung, J. Kim, Development and application of [43] C. Chen, M. Zhang, Y. Liu, S. Ma, Neural attentional rating regression with review-
machine learning-based prediction model for distillation column, Int. J. Intell. Syst. level explanations, Web Conf. 2018 - Proc. World Wide Web Conf. WWW 2018.
36 (5) (2021) 1970–1997, https://doi.org/10.1002/int.v36.510.1002/int.22368. (2018) 1583–1592. https://doi.org/10.1145/3178876.3186070.
[40] A.H. Moghaddam, M.H. Moghaddam, M. Esfandyari, Predicción del índice del [44] D.P. Kingma, J.L. Ba, Adam: A method for stochastic optimization, 3rd Int. Conf.
mercado bursátil utilizando una red neuronal artificial, J. Econ. Financ. Adm. Sci. Learn. Represent. ICLR 2015 - Conf. Track Proc. (2015) 1–15.
21 (2016) 89–93, https://doi.org/10.1016/j.jefas.2016.07.002. [45] Z. Wang, A.C. Bovik, Error : Love It or Leave It ? IEEE Signal Process. Mag. 26
[41] C. Joo, H. Park, J. Lim, H. Cho, J. Kim, Development of physical property (2009) 98–117. http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumbe
prediction models for polypropylene composites with optimizing random forest r=4775883.
hyperparameters, Int. J. Intell. Syst. (2021), https://doi.org/10.1002/int.22700. [46] T. Chai, R.R. Draxler, Root mean square error (RMSE) or mean absolute error
[42] S. Lathuiliere, P. Mesejo, X. Alameda-Pineda, R. Horaud, A comprehensive analysis (MAE)? -Arguments against avoiding RMSE in the literature, Geosci. Model Dev. 7
of deep regression, IEEE Trans. Pattern Anal. Mach. Intell. 42 (9) (2020) (2014) 1247–1250, https://doi.org/10.5194/gmd-7-1247-2014.
2065–2081, https://doi.org/10.1109/TPAMI.3410.1109/TPAMI.2019.2910523.

12

You might also like