Accepted Manuscript Process Simulation and Dynamic Control for Marine Oily Wastewater Treatment using UV Irradiation Liang Jing, Bing Chen, Baiyu Zhang, Pu Li PII:

S0043-1354(15)00205-5

DOI:

10.1016/j.watres.2015.03.023

Reference:

WR 11219

To appear in:

Water Research

Received Date: 27 October 2014 Revised Date:

23 March 2015

Accepted Date: 25 March 2015

Please cite this article as: Jing, L., Chen, B., Zhang, B., Li, P., Process Simulation and Dynamic Control for Marine Oily Wastewater Treatment using UV Irradiation, Water Research (2015), doi: 10.1016/ j.watres.2015.03.023. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

1

Process Simulation and Dynamic Control for Marine Oily Wastewater

2

Treatment using UV Irradiation

4

Liang Jing, Bing Chen*, Baiyu Zhang and Pu Li

5

RI PT

3

Northern Region Persistent Organic Pollution Control (NRPOP) Laboratory, Faculty of Engineering

7

and Applied Science, Memorial University of Newfoundland, St. John's, NL A1B 3X5, Canada

SC

6

9

M AN U

8

*Corresponding author, Tel: +1 (709) 864-8958, Fax: +1 (709) 864-4042, Email: [email protected]

AC C

EP

TE D

10

1

ACCEPTED MANUSCRIPT

11

Abstract UV irradiation and advanced oxidation processes have been recently regarded as promising

13

solutions in removing polycyclic aromatic hydrocarbons (PAHs) from marine oily wastewater.

14

However, such treatment methods are generally not sufficiently understood in terms of reaction

15

mechanisms, process simulation and process control. These deficiencies can drastically hinder their

16

application in the shipping and offshore petroleum industries which produce bilge/ballast water and

17

produced water as the main streams of marine oily wastewater. In this study, the factorial design of

18

experiment was carried out to investigate the degradation mechanism of a typical PAH, namely

19

naphthalene under UV irradiation in seawater. Based on the experimental results, a three-layer feed-

20

forward artificial neural network simulation model was developed to simulate the treatment process

21

and to forecast the removal performance. A simulation-based dynamic mixed integer nonlinear

22

programming (SDMINP) approach was then proposed to intelligently control the treatment process by

23

integrating the developed simulation model, genetic algorithm and multi-stage programming. The

24

applicability and effectiveness of the developed approach were further tested though a case study. The

25

experimental results showed that the influence of fluence rate and temperature on the removal of

26

naphthalene were greater than those of salinity and initial concentration. The developed simulation

27

model could well predict the UV-induced removal process under varying conditions. The case study

28

suggested that the SDMINP approach, with the aid of the multi-stage control strategy, was able to

29

significantly reduce treatment cost when comparing to the traditional single-stage process optimization.

30

The developed approach and its concept/framework have high potential of applicability in other

31

environmental fields where a treatment process is involved and experimentation and modeling are used

32

for process simulation and control.

AC C

EP

TE D

M AN U

SC

RI PT

12

33

2

ACCEPTED MANUSCRIPT

34

Keywords: marine oily wastewater, process simulation, process control, SDMINP, UV irradiation

35

1. Introduction

RI PT

36

Besides accidental oil spills, a disproportionately large amount of marine oil pollution is caused by

38

the regular discharge of oily wastewater from shipping and offshore oil and gas operations, such as

39

bilge water, ballast water and produced water (Jing et al., 2012a; Li et al., 2014). Many toxic and

40

persistent substances, such as polycyclic aromatic hydrocarbons (PAHs), in marine oily wastewater

41

may result in negative impacts on aquatic life and human health through food chain (Hua et al., 2007;

42

Tsapakis et al., 2010; Jing and Chen, 2011; Harman et al., 2011). Gravity settling separation (e.g.,

43

hydrocyclone) and mechanical coalescence are well-known traditional oil-water separation techniques

44

(Jing et al., 2012a); however, their efficiency largely depends on the size of oil droplets and they are

45

not effective in removing dissolved organics such as PAHs. Secondary treatment is therefore much

46

desirable after separation processes to further reduce the emission of hydrocarbons (Haritash and

47

Kaushik, 2009; Harman et al., 2011). Due to space, efficiency, cost, and/or safety concerns, many

48

traditional techniques are not readily applicable on ships or offshore oil and gas production facilities.

49

UV irradiation and its integration with other oxidation techniques (e.g., UV/O3), on the other hand,

50

have been gaining significant attention and regarded as promising alternatives for marine wastewater

51

treatment. The features of small footprint, on-site generation/operation, low cost and risk, and high

52

efficiency make them well suited for applications in remote offshore oil production platforms or

53

vessels where transportation, space, safety and cost are key concerns.

AC C

EP

TE D

M AN U

SC

37

54

Many studies have experimentally investigated the degradation of PAHs in water and wastewater

55

by using UV irradiation or its combination with other oxidants, such as ozone and H2O2 (Sanches et al.,

56

2011; Liu et al., 2014; Jing et al., 2014a). Woo et al. (2009) characterized the photocatalytic 3

ACCEPTED MANUSCRIPT

degradation efficiency and pathways of five major PAHs in aqueous solution. Kwon et al. (2009)

58

reported an inverse relationship between the reaction constant and the number of molecules for the

59

UV-induced degradation of phenanthrene and pyrene. Włodarczyk-Makuła (2011) claimed that the

60

removal efficiency of PAHs using UV was proportional to the duration and intensity of irradiation.

61

Salihoglu et al. (2012) discovered that increasing temperature would have a direct positive effect on

62

the destruction of PAHs. Tehrani-Bagha et al. (2012) applied UV-enhanced ozonation processes to

63

destruct two organic surfactants and confirmed that the synergistic effects of ozone and UV were more

64

effective than the individual processes. Liu et al. (2014) tested the removal efficiency of PAHs using

65

UV irradiation with both suspended and immobilized TiO2 catalysts. However, most of these studies

66

have focused on freshwater systems rather than marine environments where salinity and complex

67

matrix effects may play a dominant role. It is still unclear how efficient these techniques are in

68

removing PAHs from marine oily wastewater and how changing conditions (e.g., chemical dose and

69

temperature) can influence the efficiency.

TE D

M AN U

SC

RI PT

57

In addition, the research efforts on numerical modeling and performance optimization of these

71

techniques have also been limited due to their multi-physics nature and complexity of synergistic

72

effects. Many researchers have engaged in the development and application of process control and

73

optimization tools, particularly for wastewater treatment, to improve the performance and cost-

74

efficiency (Frontistis et al., 2010; Li et al., 2012; Jing et al., 2012b). Many soft computing methods,

75

including artificial neural networks, adaptive network-based fuzzy inference system, genetic algorithm,

76

and fuzzy logic, have received increasing attention and been dramatically extended to environmental

77

systems in the past decade (Elmolla et al., 2010; Mullai et al., 2011; Lin et al., 2012; Ghaedi et al.,

78

2014). These methods and their combinations can achieve higher tolerance towards imprecision,

79

uncertainty, partial truth and approximation (Han and Qiao, 2011; Bhatti et al., 2011; Ma et al., 2011;

AC C

EP

70

4

ACCEPTED MANUSCRIPT

Liu et al., 2013a; Badrnezhad and Mirza, 2014). Nonetheless, there has been limited investigation

81

reported to integrate processes simulation and dynamic control with systems optimization, and

82

especially apply them to marine wastewater treatment. Traditional optimization problems are usually

83

of static nature and assume that the values of decision variables do not change over the planning

84

horizon. However, it is not uncommon that people come across a number of situations where the

85

decision variables need to change with time. Instead of optimizing treatment systems statically with

86

fixed operation parameters, the algorithm of dynamic control can be used to rationally make a number

87

of subsequent decisions over a period of time and therefore to achieve best overall performance in

88

terms of cost or efficiency (Yu et al., 2008; Liu et al., 2010; Ferrero et al., 2012).

M AN U

SC

RI PT

80

Therefore, the objective of this study was to help fill the above gaps by integrating process

90

simulation, dynamic control and systems optimization based on experimental investigation to support

91

marine wastewater treatment. The removal mechanism of PAHs from marine oily wastewater was

92

firstly examined by UV induced photodegradation. A most commonly detected PAH in marine oily

93

wastewater, naphthalene, was targeted (Jing et al., 2014a). Based on the experimental results, an

94

artificial neural network (ANN) model was developed to simulate the removal process and predict the

95

treatment performance (Jing et al., 2014b). Finally, a simulation-based dynamic mixed integer

96

nonlinear programming (SDMINP) approach was proposed to intelligently control the process and

97

optimize the treatment system. It should be noted that the data reported in Jing et al. (2014a and 2014b)

98

were only used as a demonstrative example to examine the applicability of the SDMINP approach.

EP

AC C

99

TE D

89

100

2. Methodology

101

2.1 Experimentation

5

ACCEPTED MANUSCRIPT

Naphthalene (>99%) and naphthalene D8 (>99%, internal standard) were purchased from Aldrich,

103

Canada. Dichloromethane and acetone were purchased from Honeywell Burdick and Jackson (USA)

104

for preparing naphthalene stock solutions by 1:1 dilution (v/v). Distilled water was obtained from an

105

onsite distillation unit while seawater with a natural salinity around 25 practical salinity unit (psu) was

106

obtained from a clean coastal site in St. John’s, Canada. Filtration (with 5 µm filter) was carried out

107

prior to using the seawater in order to remove suspended particles and organisms. The photoreactor

108

used for the photodegradation experiments has a clear fused quartz beaker and an outer aluminum

109

jacket for light and heat insulation. The inner beaker features a stainless steel paddle stirring rod, a 50

110

W heater, and a thermometer, which are mounted on a polycarbonate top lid. Eight 18.4 W low-

111

pressure UV lamps (254 nm peak, Atlantic Ultraviolet, Canada) with a full width half maximum

112

(FWHM) of 15 nm are evenly mounted around the sleeve to provide adjustable irradiance.

M AN U

SC

RI PT

102

A full factorial design of experiments was employed to study the individual and synergetic effects

114

of fluence rate, salinity, temperature, and initial concentration on the removal of naphthalene from oily

115

seawater. UV fluence rate, salinity, temeperature and initial concentration were examined at 2 and 6

116

UV lamps (i.e., 2.88 and 8.27 mW cm-2, respectively), 25, 32.5, and 40 psu (i.e., average values of

117

salinity in the Newfoundland Grand Banks, the North Atlantic, and offshore produced water), 23 and

118

40 oC (i.e., room temperature and typical marine oily wastewater temperature), and 10 and 500 µg L−1

119

(i.e., typical lower and upper bounds of naphthalene concentration in offshore produced water in

120

Atlantic Canada), respectively (OGP, 2002; Blanchard et al., 2011; Neff et al., 2011; Han et al., 2011;

121

Leichsenring and Lawrence, 2011). It is worth noting that salinity was examined at three levels

122

because it is the key difference between freshwater and seawater and the presence of sodium chloride

123

can enhance the photodegradation efficiency of naphthalene in freshwater (Damle, 2008; Zheng et al.,

124

2012). During each of the 24 experiments, naphthalene stock solution was first spiked into 6 L

AC C

EP

TE D

113

6

ACCEPTED MANUSCRIPT

seawater and then vigorously stirred for 20 min to reach the thermal and volatilization equilibria. UV

126

lamps were turned on and allowed for 20 min warming-up prior to the test. A 20 ml water sample was

127

periodically collected from the reactor at a 30 min interval up to 4 hours. Ten-millilitre collected

128

sample was then transferred to a glass centrifuge tube with internal standard and 0.25 ml

129

dichloromethane for extraction. After shaking and centrifuging for 15 min each, 50 µL organic phase

130

extract was transferred into a 150 µL micro vial and analyzed by a gas chromatograph (GC) (Agilent

131

7890A) equipped with a fused silica capillary column (30 m × 0.25 mm × 0.25 µm) and a mass

132

selective detector (MS) (Agilent 5975C). The GC settings were the same as described by Jing et al.

133

(2014a). The results were analyzed by analysis of variance (ANOVA) for statistical comparisons.

M AN U

SC

RI PT

125

134 135

2.2 ANN Modeling

ANN can simulate and predict complex patterns by learning the relationships between inputs and

137

outputs. This feature allows ANN to deal with various nonlinear systems where traditional parameter

138

estimation techniques are not convenient. A typical ANN is comprised of neurons that are grouped into

139

an input layer, a hidden layer, and an output layer where the neurons are interconnected with the ones

140

on the preceding or succeeding layers by transformation equations (Jing et al., 2014b). In this research,

141

the aforementioned full factorial experimental results were used to develop a three-layer feed-forward

142

ANN with one hidden layer. As the removal rate at the beginning of the photodegradation process is

143

always zero, the first time point (time = 0) was not included. The neural network toolbox V6.0 of

144

MATLAB® was used. Experimental datasets (192 in total) were randomly divided into training (60%,

145

116 datasets), validation (20%, 38 datasets), and testing (20%, 38 datasets) subsets. The model was

146

trained by the Levenberg-Marquardt backpropagation algorithm to predict the removal rate of

147

naphthalene based on the input variables including fluence rate, salinity, temperature, initial

AC C

EP

TE D

136

7

ACCEPTED MANUSCRIPT

concentration and reaction time. The number of neurons in the hidden layer was optimized at 12. The

149

transfer functions used for the hidden and output layers were log-sigmoid (logsig) and linear (purelin),

150

respectively. The developed ANN model of naphthalene degradation was used as a demonstrative

151

example to test and validate the SDMINP approach.

RI PT

148

152

154 155

2.3 The SDMINP Approach

The proposed SDMINP approach is summarized as follows. Consider the following multi-stage

SC

153

nonlinear optimization problem in which a number of decisions have to be made on a stage-basis:
 ( ,  )

(1)

158

( ,  ) = 0,  = 1,2, … , 

(2)

159

ℎ ( ,  ) ≤ 0,  = 1,2, … , 

(3)

157

M AN U

156 subject to:

 ≤  ,  ≤ 

(4)

TE D

160

where xi and yi are the decision variables at the ith stage; i is the number of stage; f is the nonlinear

162

objective function that needs to be minimized; g and h are the nonlinear equality and inequality

163

constraints, respectively; p and q are the numbers of equality and inequality constraints, respectively;

164

and lb and ub are the lower and upper bounds of xi and yi, respectively. Most nonlinear optimization

165

problems are of static nature and assume that the values of decision variables do not change over the

166

planning horizon. However, instead of making a onetime static decision, such a multi-stage problem

167

requires a series of decisions made at each stage. The decision at each stage would have a ripple effect

168

on the succeeding decisions. Particularly, when the objective function and/or constraints are coupled

169

with simulation models, multi-stage decisions can help meet the control needs in terms of operating

170

cost, control stability, and response time.

AC C

EP

161

8

ACCEPTED MANUSCRIPT

Such types of nonlinear problems tend to possess a dynamic nature and may not be easily tackled

172

by traditional techniques. To remediate this situation, genetic algorithm (GA), which is a probabilistic

173

global optimization technique, can be employed. GA combines the “survival of the fittest” principle of

174

natural evolution with a randomized information exchange which helps to form a stochastic search

175

routine and produce new individuals with higher fitness. The detailed description of GA can be found

176

in Bhatti et al (2011) and Soleimani et al. (2013).

178

The stepwise procedure for implementing the SDMINP approach can be summarized as follows

SC

177

RI PT

171

(Fig. 1):

Step 1: Define the problem with target inputs and the number of time-stages. Set the generation

180

count Ng to zero. Set population size Np, iteration number Ni, reproduction probability pr, crossover

181

probability pc, and mutation probability pm.

M AN U

179

Step 2: Create the initial population with the preset population size (e.g., Np = 200) using random

183

binary digits. Each solution string in the population has the same number of segments as the number of

184

decision variables.

TE D

182

Step 3: Extract the values of each decision variable by reading and decoding certain sets of binary

186

digits. Apply individual solutions to the predefined simulation model(s), run the simulation for each

187

time stage i and obtain the simulation outputs.

EP

185

Step 4: Evaluate the objective function and fitness score of each solution string. Rank the strings in

189

the descending order of their fitness values. All the linear constraints and bounds should be satisfied,

190

while the nonlinear constraints may not be all satisfied at every generation but will be met at the

191

convergence solution.

AC C

188

9

ACCEPTED MANUSCRIPT

192

Step 5: Create a mating pool by using a proportional selection criterion (e.g., simulated roulette) to

193

choose solutions in the initial population. The probability of an individual solution being chosen is

194

proportional to its fitness value.

196

Step 6: From the mating pool, choose parent strings to perform reproduction (Pr) and crossover (Pc) operations to obtain the offspring population.

RI PT

195

Step 7: Perform mutation (Pm) on the offspring population.

198

Step 8: Update the generation count by Ng = Ng + 1.

199

Step 9: Repeat steps 3-8 to update the population pool with new generations until one of the

200

convergence criteria is satisfied, such as the generation count Ng reaches the preset limit, or the

201

weighted average relative change in the fitness function value over stall generations is less than the

202

function tolerance. The convergence solution is record as the final solution to the optimization problem.

M AN U

203

2.4 Case Study: Marine Oily Wastewater Treatment

TE D

204

SC

197

Consider the following marine oily wastewater treatment system to examine the effectiveness of

206

the SDMINP approach in optimizing operation strategy. The flow-through treatment system consists of

207

a storage tank and a reaction tank with maximum capacities of 15 and 3 L, respectively (Fig. 2). The

208

removal of naphthalene from oil polluted seawater by using UV irradiation is used as an example. The

209

removal process occurs only in the reaction tank, where the average UV fluence rate is assumed to be

210

the same as described in Section 2.1. Two pumps with adjustable flow rates are used to circulate

211

seawater through the treatment system. Controlled variables include the number of UV lamps, pump

212

flow rate, initial naphthalene concentration, salinity, temperature and discharge standard.

AC C

EP

205

213

A total volume of 15 L seawater polluted by naphthalene at a concentration of 300 µg L−1 needs to

214

be treated using this system. The concentration is determined based on the average value of the

10

ACCEPTED MANUSCRIPT

literature data and the sample measurements (OGP, 2002; Blanchard et al., 2011; Jing et al., 2014a).

216

More detailed information about the concentration ranges of NAP and other PAHs in offshore

217

produced water can be found in the literature (Neff et al., 2011; Jing et al., 2012a; Liu et al., 2014;

218

Chen et al., 2014). It is assumed that, while circulating, 12 and 3 L seawater is retained in the storage

219

and reaction tanks, respectively. Salinity and water temperature are maintained at 32.5 practical

220

salinity unit (psu) and 25 oC, respectively. According to the stringent marine water quality standard

221

(CCME, 1999; McLaughlin et al., 2014), the concentrations of naphthalene in both tanks need to be

222

lower than 10 µg L−1 prior to discharge. The objective of this case study is therefore to design an

223

hourly operation strategy for 36 hours in order to meet the discharge standard and to minimize the total

224

treatment cost. Cost is assumed to be associated with the use of UV lamps and pumps by which

225

electricity is consumed. Their cost coefficients are set as $0.05 per lamp per hour and $0.003 per liter

226

water pumped, respectively. UV lamps can be operated in different combinations and the number has

227

to be an integer ranging from 2 to 6. The flow rates of the pumps are set as equal and can vary from 0.1

228

to 0.5 L min-1. This problem can then be formulated as the following multi-stage mixed integer

229

nonlinear programming problem with 36 stages (one hour per stage).

TE D

M AN U

SC

RI PT

215

232 233 234 235

subject to:

AC C

231


 (, ) = ∑$% &'(0.003 × 60 + 0.05  )

(5)

()*+,-./ (' , 0 , … , $% , ' , 0 , … , $% ) ≤ 10

(6)

(,/-1*+2 (' , 0 , … , $% , ' , 0 , … , $% ) ≤ 10

(7)

EP

230

0.1 ≤  ≤ 0.5  = 1,2, … ,36

(8)

2 ≤  ≤ 6 and integer  = 1,2, … ,36

(9)

236

where variables xi and yi are the flow rate (L min-1) and the number of UV lamps at the ith stage,

237

respectively; and C are the nonlinear inequity constraints on the concentrations of naphthalene after

11

ACCEPTED MANUSCRIPT

238

36-hour treatment (µg L−1). The computation of C is simplified as follows. Take the first hour as an

239

example, after a small time step ∆t , the concentrations in the storage (cs) and reaction (cr) tanks

240

(assuming completely mixed conditions) can be computed as:

242

1) The removal rate of naphthalene (r∆t) in the portion of seawater that remains in the reaction tank

RI PT

241

(SWr) after ∆t can be calculated by the developed ANN simulation model

;∆* = sim(net, ?@,A ; 32.5; ' ; 25; ∆CD)

243

(14)

where sim and net are Matlab commands for ANN simulation; y1 is the number of UV lamps during

245

the first hour; cr0, 32.5, and 25 stand for the concentration in the reaction tank prior to ∆t, salinity, and

246

temperature, respectively.

M AN U

SC

244

247

2) Within ∆t, the volume of seawater that flows out from the reaction tank (SWout) is x1∆t. As this

248

portion gradually flows out and receives different amount of irradiation, the overall removal rate rout

249

can be integrated as shown below. According to the first order reaction kinetics, we have ln

TE D

250

1F∆G 1FH

;∆* = 1 −

251

;* = 1 −

EP

252

1FH

1F∆G 1FH

= 1 − J K∆*

= 1 − J K* = 1 − J '

(10)

LMNOPF∆G QG ∆G

(11)

∆*

;+R* = SA ;* TC *

(12) (13)

AC C

253

1FG

= −∆C

254

where x1 is the flow rate during the first hour; k is the photoreaction rate constant; cr∆t and crt are the

255

concentrations in SWr after ∆t and at time point t, respectively; and rt is the corresponding removal rate

256

at time point t.

257 258 259

3) Within ∆t, the volume of seawater that flows into the reaction tank (SWin) is also x1∆t and the overall removal rate rin can be integrated using the same procedure stated above. ∗ ;∆* = sim(net, ?@)A ; 32.5; ' ; 25; ∆CD)

12

(14)

ACCEPTED MANUSCRIPT

;*∗

260

∆*

'

(15) (16)

where cs0 is the concentration of naphthalene in the storage tank prior to ∆t; r∆*t and rt* are the

RI PT

263



;2 = * SA ;*∗ TC

261 262

=1−J

LMNOPF∗∆G QG ∆G

removal rates in seawater coming from the storage tank at time points ∆t and t, respectively.

4) Seawater is assumed to be completely mixed in both tanks after ∆t. The total naphthalene

265

removal in mass m and the concentrations in the storage (cs) and reaction tanks (cr) after ∆t can be

266

computed as:

SC

264

@, =

268

M AN U

V = (3 − ' ∆C)@,A ;∆* + ' ∆C@,A ;+R* + ' ∆C@)A ;2

267

∗ ($KWO ∆*)1FH ('K,∆G )XWO ∆*1YH ('K,∆G )

@) =

269

$

('0KWO ∆*)1YH XWO ∆*1FH ('K,∆G ) '0

(17) (18) (19)

Repeat the above four steps for the rest of the first hour as well as subsequent hours to obtain the

271

final concentrations in both tanks, which are required to be less than or equal to 10 µg L−1. It can be

272

seen that the length of ∆t is of significance in computing the concentrations of naphthalene in both

273

tanks, which are in turn affected by the differences among the removal rates r∆t, rout, and rin. Ideally, ∆t

274

should be set short enough to approach the complete mixing assumption such that the differences

275

among r∆t, rout, and rin are negligible. Therefore, in this case study, ∆t was tested at 0.1, 0.5, 1, 2, 3, 4,

276

and 5 min, respectively, with cr0 varying from 300 to 100 µg L-1 to determine the best compromise

277

between computation accuracy and efficiency. As shown in Table 1, the concentration difference (∆c)

278

between r∆t and rout/rin shows an increasing trend when ∆t increases and this trend is not affected by cr0.

279

However, to simplify the computation and avoid time-consuming iterations, a compromise on the

280

length of ∆t should be taken into account such that ∆t was set as 1 min in this case study.

AC C

EP

TE D

270

13

ACCEPTED MANUSCRIPT

To solve this problem, GA optimization was performed in Matlab® by following the detailed

282

procedure shown in Fig. 1. The number of variables was 72. Population size (Np) and maximum

283

generation count (Ng) are usually two of the most important paramters in GA optimization. To validate

284

if the chosen values of Np and Ng can well achieve the desired solution, a comparative test was

285

undertaken by examining their different combinations (Table 2). It can be seen that the optimization

286

results dropped from $9.6784 to $9.1125 when Np was held at 30 and Ng was increased from 50 to 100.

287

However, a further increase of Ng from 100 to 200 did not produce any better results because the

288

optimization process usually met the stopping criteria between 70 to 80 generations. On the other hand,

289

higher Np values were not associated with better performance as observed from the last three tests. It

290

has been reported that big population size may not improve the performance of GA in sense of speed

291

of finding solution (El-Fergany et al., 2014). Np was therefore set to 30 by taking computation time and

292

resource constraints into account. Based on the test results and literature recommendations (Ma et al.,

293

2011; Dong et al., 2013; Liu et al., 2013b), the population size (Np) and maximum generation count (Ng)

294

were set at 30 and 100, respectively. The binary tournament selection function and doubleVector

295

population type were used as default for mixed integer programming. Binary tournament selection can

296

be explained as the direct pairwise comparisons among the individuals in order to choose the most

297

successful one, while doubleVector specifies that the type of individuals is double. The elite count for

298

reproduction (Pr) was set at 2, while the crossover fraction (Pc) of the current population excluding the

299

elite individuals was set at 0.8. The power mutation (Pm) was adopted by default to accommodate the

300

integer restriction of decision variables (Deep et al., 2009).

AC C

EP

TE D

M AN U

SC

RI PT

281

301

To compare with the performance of the SDMINP approach, a single-stage optimization (i.e.,

302

continuous optimization problem) was conducted by using constant flow rate and constant number of

303

lamps throughout the 36-hour period. The optimization procedure and GA settings were the same as

14

ACCEPTED MANUSCRIPT

for the multi-stage problem. In addition, a 500-iteration Monte Carlo simulation was performed to

305

illustrate the distribution of treatment cost that might be expected from random sampling of constant

306

flow rate and the number of lamps within their ranges. Such a random sampling can well reflect the

307

consequence of random decision making when no optimization results are available.

RI PT

304

309

3. Results and Discussion

310

3.1 Experimentation and ANN Modeling

SC

308

Based on the experimental results (Fig. 3), it can be seen that all input parameters including initial

312

concentration, salinity, UV fluence rate, and temperature had noticeable effects on the removal of

313

naphthalene. Statistical results from ANOVA suggested that all four factors had significant influence

314

with a p-value less than 0.05. However, the contribution of initial concentration and salinity was not as

315

remarkable as that of fluence rate and temperature. Fig. 3a and 3b show that initial concentration and

316

salinity were not the most significant contributors to the reaction mechanism as the lines are close to

317

each other. Nonetheless, increasing salinity from 25 to 32.5 and 40 psu could to some extent suppress

318

the degradation of naphthalene. This may be explained by the presence and increase of free radical

319

scavengers such as bromide (Br−), carbonate (CO32−) and bicarbonate (HCO3−) ions (Lair et al., 2008).

320

When initial concentration varying from 10 to 500 µg L−1, it was observed that the reaction rate

321

constant was slightly lowered from 0.0183 to 0.0176 min-1, possibly due to the lack of reactive radicals.

322

Contrastingly, as illustrated in Fig. 3c and 3d, increasing fluence rate from 2.88 (i.e., 2 lamps) to

323

8.27 mW cm-2 (i.e., 6 lamps) could drastically enhance the photodegradation reaction rate constant by

324

up to 200%. Such an increase was equally remarkable in both experimental and modeling results. This

325

was attributed to the fact that increasing the fluence rate would promote the collision between activable

326

centers and therefore create more oxidative radicals (e.g., hydroxyl radical OH·). Such a positive trend

AC C

EP

TE D

M AN U

311

15

ACCEPTED MANUSCRIPT

was also noted for temperature as higher temperature is able to increase the collision between photons

328

and target molecules. Fig. 3d depicts a comparison between the modeling and experimental

329

degradation results at 23 and 40 oC. It can be seen that the removal performance at 40 oC was

330

constantly superior to that at 20 oC, regardless of time. It can be further observed that the ANN

331

predicted results were in good agreement with the experimental results. The coefficients of

332

determination (R2) between all the experiment and modeling results (e.g., solid and open circles in Fig.

333

3a) were greater than 0.95. Such a good fit showed that the developed ANN model was capable of

334

accurately simulating the PAH removal process and therefore served as the foundation of the proposed

335

SDMINP approach. However, it is noteworthy to mention that due to ANNs’ generality issues, a re-

336

training process would be necessary if the ranges of the parameters or the model response (e.g., target

337

pollutant) were changed or the proposed method was used for other cases.

M AN U

SC

RI PT

327

338

3.2 Optimization of the Treatment System

TE D

339

The results from the case study suggested that the minimum treatment cost was optimized at

341

$9.1125 after 36-hour UV irradiation. The final concentrations in the storage and reaction tanks were

342

9.97 and 7.56 µg L−1, respectively, indicating that the nonlinear constraints on the discharge standards

343

were met (Fig. 4). The decreasing trend of both concentrations were closely linked with a correlation

344

coefficient of 0.998, while they were negatively correlated with treatment cost, with correlation

345

coefficients lower than -0.938. This observation suggests that the decision variables were properly

346

optimized. If the flow rate was too low, given that the number of UV lamps remained unchanged, the

347

concentration difference between two tanks would be more pronounced because most treated seawater

348

would remain in the reaction tank. If the flow rate was too high, the pollutant removal process could be

349

somewhat accelerated but with much higher cost. On the other hand, given a constant flow rate, more

AC C

EP

340

16

ACCEPTED MANUSCRIPT

UV lamps may not only drastically accelerate the treatment process but also associate with increasing

351

cost, and vice versa. The flow rates and the number of UV lamps during each stage are illustrated in

352

Fig. 5. It is worth mentioning that the treatment system was not at its peak load condition. The number

353

of UV lamps was kept at its minimum level (e.g., 2 lamps) in 16 hours, and the mean flow rate during

354

the whole treatment period was just half of its upper limit at 0.241 L min-1. This implies that the 36-

355

hour period may be further shortened without violating the discharge standards.

RI PT

350

357

SC

356

3.3 Comparison with Single-stage Optimization and Monte Carlo Simulation The single-stage optimization yielded a constant flow rate of 0.1 L min-1 and a constant UV fluence

359

rate of 6 lamps during the 36-hour period. The variations of naphthalene concentrations in both tanks

360

with time are plotted in Fig. 6. The degradation strictly followed the first order kinetics well with R2

361

greater than 0.99 and reduced the concentrations in the storage and reaction tanks to 4.15 and 3.08 µg

362

L−1, respectively. The difference between two concentrations ranged from 1.07 to 45.92 µg L−1 with a

363

mean value of 15.75 µg L−1. The mean value was larger than that of the multi-stage problem (i.e., 6.53

364

µg L−1), indicating that the fixed parameters were not as efficient as the variable ones in balancing the

365

concentrations. The treatment cost was optimized to $11.448, which was 25.7% higher than that of the

366

multi-stage problem, suggesting that the multi-stage operation plan was superior in terms of cost

367

efficiency. This difference could be solely due to the over-exposure of UV irradiation by using 6 lamps

368

during the treatment period as the flow rate was at its minimum level (0.1 L min-1). If the number of

369

UV lamps was adjusted to 5 while maintaining the flow rate at 0.1 L min-1, the concentrations in the

370

storage and reaction tanks went down to 16.52 and 13.55 µg L−1, respectively, which were above the

371

discharge standard. The standard was not met until the flow rate was increased to 0.5 L min-1 with 5

372

lamps on, incurring a total cost of $12.24. Therefore, using the combination of 0.1 L min-1 flow rate

AC C

EP

TE D

M AN U

358

17

ACCEPTED MANUSCRIPT

and 6 UV lamps seemed to be the best decision available for this single-stage treatment plan. Another

374

interesting finding was that the concentrations in both tanks were lower than 10 µg L−1 after only 29

375

hours, indicating the rest of the treatment period could be readily considered as unnecessary. Therefore,

376

as concluded from the multi-stage problem, a reduction in treatment duration could be favoured

377

because it may also lead to lower treatment cost.

RI PT

373

On the other hand, 83% (415 out of 500) of the Monte Carlo iterations were not able to meet the

379

required standard (10 µg L−1) due to the lack of flow rate or UV irradiation. Only 85 iterations were

380

recorded as feasible solution and the statistical analysis of treatment cost, flow rate, and the number of

381

lamps is shown in Table 3. It can be seen that the minimum treatment cost was $11.449, which was

382

close to the single-stage optimization results of $11.448. Nonetheless, the mean cost was $12.797 with

383

a standard deviation of $0.728, indicating that in most Monte Carlo iterations the system performance

384

was not at its optimal level. These results suggested that, if the operator randomly set the flow rate and

385

the number of lamps as constants during the 36-hour period, then there would be a great chance that

386

the treatment standard cannot be met. Even if the treatment was successful, the system may hardly

387

reach its optimal performance due to the lack of optimization efforts.

389

M AN U

TE D

EP

388

SC

378

3.4 Minimization of Time and Cost

The original problem of the case study was to minimize treatment cost given a 36-hour time period.

391

The optimization results showed that the treatment system was not operated at full capacity, indicating

392

the operation time may be reduced. The comparative single-stage problem also revealed that a

393

reduction in treatment duration could be necessary and cost-effective. Indeed, increasing flow rate and

394

UV irradiation intensity may help achieve the discharge standard in a shorter operating time. To better

395

understand the relationship between cost and operation time, the SDMINP approach and the single-

AC C

390

18

ACCEPTED MANUSCRIPT

stage optimization were both carried out for multiple periods ranging from 25 to 36 hours by taking

397

time constraints and the efficiency of convergence into account. The number of stages of the SDMINP

398

approach was set the same as the number of hours (one hour per stage) for each period. The validity of

399

constraints (final concentrations not exceeding 10 µg L−1) was checked such that feasible solutions in

400

terms of cost and time were recorded. The population size (Np), maximum generation count (Ng) and

401

elite count for reproduction (Pr) were kept as 30, 100 and 2, respectively.

RI PT

396

As depicted in Fig. 7, it can be seen that the treatment costs in both multi-stage SDMINP and

403

single-stage optimization were reduced with decreasing time. This decreasing trend was remarkable at

404

the beginning from 36 hours and reached a local minimum around 27 hours for both cases. The costs

405

started to increase if the treatment time went below 27 hours and reached the threshold at 25 hours for

406

the discharge standard to be met. The discharge standard of 10 µg L−1 was not able to be satisfied if the

407

treatment time was less than 25 hours. Another interest finding is that the optimum cost of the multi-

408

stage control scheme was lower than that of the single-stage control scheme at an average difference of

409

$1.00, implying the superiority of the SDMINP approach.

M AN U

TE D

411

3.5 Effect of the Number of Optimization Stages

EP

410

SC

402

The SDMINP approach was originally proposed and implemented on an hourly basis so that each

413

stage had one hour. However, such an hourly operation plan sometimes may not be the best option,

414

given that changing the operation parameters too often may cause additional costs or affect the stability

415

of the system. To understand how the number of optimization stages can influence the optimization

416

results, sensitivity analysis was conducted. When the total treatment period was fixed at 36 hours, it

417

can be seen that more optimization stages generally resulted in lower total treatment cost (Table 4).

418

Interestingly, when having two stages, or in other words, changing the operation parameters once after

AC C

412

19

ACCEPTED MANUSCRIPT

the 18th hour, the total treatment cost would reach its minimum value at $8.3622. The flow rates and

420

the numbers of lamps at the two stages were 0.160 and 0.199 L min-1, and 2 and 6, respectively. Such

421

an interesting observation may be explained by the nature of GA. The solution search mechanism of

422

GA often starts from a random initial population such that the 36-stage optimization case would hardly

423

converge to the solution of the 2-stage one due to the presence of more decision variables. With that

424

being said, if the solution of the 2-stage case can be used as the initial population for the 36-stage one,

425

then a better solution may be found. Their performance was further compared at a number of other

426

periods (Table 5) such that the results obtained from the hourly strategy (i.e., 36-stage) dominated

427

those from the 2-stage one, indicating the advantage of using more stages. Nonetheless, it should be

428

taken into account that increasing the number of stages would possibly increase manpower needs and

429

reduce system stability. Therefore, for the sake of computation time, the operators are recommended to

430

first seek the best solution with less optimization stages, and then use it as an initial population to seek

431

potentially better solution with more optimization stages. Then the difference between the solutions

432

can be evaluated to determine whether or not a more complex operation plan should be implemented.

434

4. Conclusions

EP

433

TE D

M AN U

SC

RI PT

419

UV irradiation and advanced oxidation techniques have been recently found to be successful for the

436

abatement of PAHs in marine oily wastewater. However, the lack of process understanding,

437

performance prediction, and process control has substantially hindered their applications. In response

438

to such gaps, this study presented a novel experiment and modeling coupling approach by integrating

439

process simulation, dynamic control and systems optimization based on experimental investigation.

440

The UV-induced photodegradation of naphthalene in oily seawater was first conducted and provided as

441

an example to examine the effectiveness of the proposed approach. The experimental results indicated

AC C

435

20

ACCEPTED MANUSCRIPT

that UV fluence rate and temperature were the most influential variables due to increased number of

443

oxidative radicals and more intensive collisions between photons and target molecules. Increasing

444

salinity would impede the removal process because of the radical scavengers like bromide (Br−),

445

carbonate (CO32−) and bicarbonate (HCO3−). A three-layer feed-forward artificial neural network

446

(ANN) simulation model was successfully developed to predict the removal performance with good

447

overall agreement. The coefficients of determination (R2) between all groups of experiment and

448

modeling results were greater than 0.95. The SDMINP approach was then developed based on the

449

simulation model, genetic algorithm and multi-stage programming.

SC

RI PT

442

A case study related to the removal of naphthalene from oil polluted seawater using UV irradiation

451

and a continuous treatment system was conducted. The results from the case study showed that the

452

treatment cost in a fixed 36-hour period was minimized to $9.11 by using the SDMINP approach. As a

453

comparison, the single-stage optimization with fixed variables was applied and the treatment cost was

454

25.7% higher at $11.45. A Monte Carlo simulation was also performed to conclude that if the operator

455

randomly set the flow rate and the number of lamps as constants during the 36-hour period, then there

456

would be a great chance of failure in meeting the treatment standard. If considering time as another

457

flexible variable, the treatment costs reached their minimum in 27 hours at $8.71 and $8.94 for the

458

SDMINP approach and the single-stage optimization, respectively, further indicating the advantage of

459

the former. A sensitivity analysis study on the number of stages demonstrated that, regardless the

460

length of treatment period, an increase in the number of optimization stages could generally reduce the

461

treatment cost, but might lead to extra manpower needs and affect system stability. It was

462

recommended to first seek the best solution with less optimization stages, and then use the solution as

463

an initial population for more optimization stages, if necessary. Ongoing studies at the NRPOP Lab

464

focus on the treatment of other parents and alkyl PAHs in oily wastewater such as onshore/offshore

AC C

EP

TE D

M AN U

450

21

ACCEPTED MANUSCRIPT

produced water and bilge water. The general concept and modeling framework that includes both

466

experimentation and the SDMINP approach can be applied with certain modifications (e.g., training of

467

ANNs for different data sets or change of simulation models) to other wastewater treatment processes

468

or any other environmental mitigation systems, where numerical models can be structured for

469

simulation and control and experiments can be conducted for data acquisition and model validation.

RI PT

465

470

Acknowledgements

SC

471

Special thanks go to Natural Sciences and Engineering Research Council of Canada (NSERC),

473

Research & Development Corporation of Newfoundland and Labrador (RDC NL), and Canada

474

Foundation for Innovation (CFI) for financing and supporting this research.

M AN U

472

475 476

References

Badrnezhad, R., Mizra, B., 2014. Modeling and optimization of cross-flow ultrafiltration using

478

hybrid neural network-genetic algorithm approach. Journal of Industrial and Engineering Chemistry,

479

20(2), 528-543.

TE D

477

Bhatti, M.S., Kapoor, D., Kalia, R.K., Reddy, A.S., Thukral, A.K., 2011. RSM and ANN modeling

481

for electrocoagulation of copper from simulated wastewater: Multi objective optimization using

482

genetic algorithm approach. Desalination, 274(1-3), 74-80.

AC C

EP

480

483

Blanchard, A.L., Feder, H.M., Shaw, D.G., 2011. Associations between macrofauna and sediment

484

hydrocarbons from treated ballast water effluent at a marine oil terminal in Port Valdez, Alaska.

485

Environmental Monitoring and Assessment, 178(1–4), 461–476.

486

CCME (Canadian Council of Ministers of the Environment), 1999. Canadian Water Quality

487

Guidelines for the Protection of Aquatic Life: Polycyclic aromatic hydrocarbons (PAHs). In: Canadian 22

ACCEPTED MANUSCRIPT

488

Environmental Quality Guidelines, Canadian Council of Ministers of the Environment, Winnipeg,

489

Canada. Chen, B., Zhang, B.Y., Husain, T., Zheng, J.S., Ma, Y.C., Liu, B., Li, Z.L., 2014. Ozonation as a

491

treatment option for produced water effluents. Technical Report, Prepared for Petroleum Research

492

Newfoundland and Labrador (PRNL), November, 251 pages.

494

Damle, N.S., 2008. Fog-smog reactor and photooxidation of naphthalene within the fog condensate in a UV light setup. M.S. Thesis, Louisiana State University, Baton Rouge, LA, U.S.

SC

493

RI PT

490

Deep, K., Singh, K.P., Kansal, M.L., Mohan, C., 2009. A real coded genetic algorithm for solving

496

integer and mixed integer optimization problems. Applied Mathematics and Computation, 212, 505-

497

518.

M AN U

495

Dong, H.L., Huang, J.J., Cai, Z.H., Wu, Q.F., 2013. Research on Predicted Model of Least Squares

499

Support Vector Machine Based on Genetic Algorithm. Advanced Materials Research, 753, 2875-2881.

500

Elmolla, E.S., Chaudhuri, M., Eltoukhy, M.M., 2010. The use of artificial neural network (ANN)

501

for modeling of COD removal from antibiotic aqueous solution by the Fenton process. Journal of

502

Hazardous Materials, 179, 127–134.

TE D

498

El-Fergany, A.A., Othman, A.M., El-Arini, M.M., 2014. Synergy of a genetic algorithm and

504

simulated annealing to maximize real power loss reductions in transmission networks. International

505

Journal of Electrical Power & Energy Systems, 56, 307-315.

507

AC C

506

EP

503

Ferrero, G., Rodríguez-Roda, I., Comas, J., 2012. Automatic control systems for submerged membrane bioreactors: A state-of-the-art review. Water Research, 46(11), 3421-3433.

508

Frontistis, Z., Daskalaki, V.M., Hapeshi, E., Drosou, C., Fatta-Kassinos, D., Xekoukoulotakis, N.P.,

509

Mantzavinos, D., 2010. Photocatalytic (UV-A/TiO2) degradation of 17α-ethynylestradiol in

23

ACCEPTED MANUSCRIPT

510

environmental matrices: Experimental studies and artificial neural network modeling. Journal of

511

Photochemistry and Photobiology A: Chemistry, 240, 33–41. Ghaedi, M., Zeinali, N., Ghaedi, A.M., Teimuori, M., Tashkhourian, J., 2014. Artificial neural

513

network-genetic algorithm based optimization for the adsorption of methylene blue and brilliant green

514

from aqueous solution by graphite oxide nanoparticle. Spectrochimica Acta Part A: Molecular and

515

Biomolecular Spectroscopy, 125, 264-277.

518 519 520 521

SC

517

Han, G.Q., Ma, Z.M., deYoung, B., Foreman, M., Chen, N., 2011. Simulation of three-dimensional circulation and hydrography over the Grand Banks of Newfoundland. Ocean Modelling, 40, 199–210. Han, H.G., Qiao, J.F., 2011. Adaptive dissolved oxygen control based on dynamic structure neural

M AN U

516

RI PT

512

network. Applied Soft Computing, 11(4), 3812-3820.

Haritash, A.K., Kaushik, C.P., 2009. Biodegradation aspects of polycyclic aromatic hydrocarbons (PAHs): a review. Journal of Hazardous Materials, 169, 1–15.

Harman, C., Brooks, S., Sundt, R.C., Meier, S., Grung, M., 2011. Field comparison of passive

523

sampling and biological approaches for measuring exposure to PAH and alkylphenols from offshore

524

produced water discharges. Marine Pollution Bulletin, 63(5–12), 141–148.

TE D

522

Hua, F.L., Tsang, Y.F., Wang, Y.J., Chan, S.Y., Chua, H., Sin, S.N., 2007. Performance study of

526

ceramic microfiltration membrane for oily wastewater treatment. Chemical Engineering Journal,

527

128(2-3), 169-175.

529 530 531

AC C

528

EP

525

Jing, L., Chen, B., 2011. Field Investigation and Hydrological Modelling of a Subarctic Wetland the Deer River Watershed. Journal of Environmental Informatics, 17(1), 36-45. Jing, L., Chen, B., Zhang, B.Y., Peng, H.X., 2012a. A review of ballast water management practices and challenges in harsh and Arctic environments. Environmental Reviews, 20, 83–108.

24

ACCEPTED MANUSCRIPT

532

Jing, L., Chen, B., Zhang, B.Y., Li, P., 2012b. A stochastic simulation-based hybrid interval fuzzy

533

programming approach for optimizing the treatment of recovered oily water. Journal of Ocean

534

Technology, 7(4), 59–72. Jing, L., Chen, B., Zhang, B.Y., Zheng, J.S., Liu, B., 2014a. Naphthalene degradation in seawater

536

by UV irradiation: the effects of fluence rate, salinity, temperature and initial concentration. Marine

537

Pollution Bulletin, 81, 149-156.

539

Jing, L., Chen, B., Zhang, B.Y., 2014b. Modeling of UV-induced photodegradation of naphthalene

SC

538

RI PT

535

in marine oily wastewater by artificial neural networks. Water, Air, & Soil Pollution, 225(4), 1-14. Kwon, S.H., Kim, J.H., Cho, D., 2009. An analysis method for degradation kinetics of lowly

541

concentrated PAH solutions under UV Light and ultrasonication. Journal of Industrial and Engineering

542

Chemistry, 15, 157–162.

M AN U

540

Lair, A., Ferronato, C., Chovelon, J.-M., Herrmann, J.-M., 2008. Naphthalene degradation in water

544

by heterogeneous photocatalysis: an investigation of the influence of inorganic anions. Journal of

545

Photochemistry and Photobiology A: Chemistry, 193, 193–203.

547 548

Leichsenring, J., Lawrence, J., 2011. Effect of mid-oceanic ballast water exchange on virus-like particle abundance during two trans-Pacific voyages. Marine Pollution Bulletin, 62(5), 1103–1108.

EP

546

TE D

543

Li, P., Chen, B., Zhang, B.Y., Jing, L., Zheng, J.S., 2012. A multiple-stage simulation-based mixed integer nonlinear programming approach for supporting offshore oil spill recovery with weathering

550

processes. Journal of Ocean Technology, 7(4), 87–105.

551

AC C

549

Li, P., Chen, B., Zhang, B.Y., Jing, L., Zheng, J.S., 2014. Monte Carlo simulation-based dynamic

552

mixed integer nonlinear programming for supporting oil recovery and devices allocation during

553

offshore oil spill responses. Ocean & Coastal Management, 89, 58-70.

25

ACCEPTED MANUSCRIPT

554

Lin, C.H., Yu, R.F., Cheng, W.P., Liu, C.R., 2012. Monitoring and control of UV and UV-TiO2

555

disinfections for municipal wastewater reclamation using artificial neural networks. Journal of

556

Hazardous Materials, 209–210, 348–354. Liu, Y.C., Shi, H.C., Shi, H.M., Wang, Z.Q., 2010. Study on a discrete-time dynamic control

558

model to enhance nitrogen removal with fluctuation of influent in oxidation ditches. Water Research,

559

44(18), 5150-5157.

RI PT

557

Liu, H.B., Huang, M.Z., Yoo, C.K., 2013a. A fuzzy neural network-based soft sensor for modeling

561

nutrient removal mechanism in a full-scale wastewater treatment system. Desalination and Water

562

Treatment, 51(31-33), 6184-6193.

M AN U

SC

560

Liu, Z., Ma, S., Shi, Y., Teng, H., 2013b. Solving multi-objective Flexible Job Shop Scheduling

564

with transportation constraints using a micro artificial bee colony algorithm. In: Proceedings of the

565

2013 IEEE 17th International Conference on Computer Supported Cooperative Work in Design, June

566

27-29, Whistler, BC, Canada, pp. 427-432.

TE D

563

Liu, B., Chen, B., Zhang, B.Y., Ma, Y.C., Jing, L., 2014. Photocatalysis of Naphthalene in offshore

568

produced water: comparison between suspended TiO2 and immobilized TiO2. In: Proceedings of the

569

2014 International Conference on Marine and Freshwater Environments, August 6-8, St. John’s,

570

Canada.

EP

567

Ma, Y.W., Huang, M.Z., Wan, J.Q., Hu, K., Wang, Y., Zhang, H.P., 2011. Hybrid artificial neural

572

network genetic algorithm technique for modeling chemical oxygen demand removal in anoxic/oxic

573

process. Journal of Environmental Science and Health Part A, 46, 574-580.

574 575

AC C

571

McLaughlin, C., Falatko, D., Danesi, R., Albert, R., 2014. Characterizing shipboard bilgewater effluent before and after treatment. Environmental Science and Pollution Research, 21(8), 5637-5652.

26

ACCEPTED MANUSCRIPT

576

Mullai, P., Arulselvi, S., Ngo, H.H., Sabarathinam, P.L., 2011. Experiments and ANFIS modeling

577

for the biodegradation of penicillin-G wastewater using anaerobic hybrid reactor. Bioresource

578

Technology, 102, 5492-5497.

581 582

In: Produced Water, Springer New York, New York, U.S.

RI PT

580

Neff, J., Lee, K., DeBlois, E.M., 2011. Produced water: overview of composition, fates, and effects.

OGP (International Association of Oil and Gas Producers), 2002. Aromatics in produced water: occurrence, fate & effects and treatment. OGP Report no. 1.20/324. London, England, 30 pp.

SC

579

Salihoglu, N.K., Karaca, G., Salihoglu, G., Tasdemir, Y., 2012. Removal of polycyclic aromatic

584

hydrocarbons from municipal sludge using UV light. Desalination and Water Treatment, 44(1–3), 324–

585

333.

M AN U

583

Sanches, S., Leitão, C., Penetra, A., Cardoso, V.V., Ferreira, E., Benoliel, M.J., Crespo, M.T.B.,

587

Pereira, V.J., 2011. Direct photolysis of polycyclic aromatic hydrocarbons in drinking water sources.

588

Journal of Hazardous Materials, 192(3), 1458–1465.

TE D

586

Soleimani, R., Shoushtari, N.A., Mirza, B., Salahi, A., 2013. Experimental investigation, modeling

590

and optimization of membrane separation using artificial neural network and multi-objective

591

optimization using genetic algorithm. Chemical Engineering Research and Design, 91(5), 883-903.

593

Tehrani-Bagha, A.R., Nikkar, H., Menger, F.M., Holmberg, K., 2012. Degradation of Two Persistent Surfactants by UV-Enhanced Ozonation. Journal of Surfactant and Detergent, 15, 59-66.

AC C

592

EP

589

594

Tsapakis, M., Dakanali, E., Stephanou, E.G., Karakassis, I., 2010. PAHs and n-alkanes in

595

Mediterranean coastal marine sediments: aquaculture as a significant point source. Journal of

596

Environmental Monitoring, 12, 958–963.

27

ACCEPTED MANUSCRIPT

597

Włodarczyk-Makuła, M., 2011. Application of UV-rays in removal of polycyclic aromatic

598

hydrocarbons from treated wastewater. Journal of Environmental Science and Health, Part A:

599

Toxic/Hazardous Substances and Environmental Engineering, 46(3), 248–257. Woo, O.T., Chung, W.K., Wong, K.H., Chow, A.T., Wong, P.K., 2009. Photocatalytic oxidation of

601

polycyclic aromatic hydrocarbons: intermediates identification and toxicity testing. Journal of

602

Hazardous Materials, 168, 1192–1199.

RI PT

600

Yu, R.F., Chen, H.W., Cheng, W.P., Shen, Y.C., 2008. Dynamic control of disinfection for

604

wastewater reuse applying ORP/pH monitoring and artificial neural networks. Resources,

605

Conservation, and Recycling, 52(8-9), 1015-1021.

M AN U

SC

603

Zheng, J.S., Liu, B., Ping, J., Zhang, B.Y., Chen, B., 2012. Two analytical methods for real-time

607

monitoring of polycyclic aromatic hydrocarbons in oil contaminated seawater. In: proceedings of the

608

35th Arctic and Marine Oilspill Program (AMOP) Technical Seminar on Environmental

609

Contamination and Response, June 5-7, Vancouver, Canada. p. 448-459.

EP AC C

610

TE D

606

28

ACCEPTED MANUSCRIPT

Table 1. The removal rates r∆t, rout, and rin after one time step ∆t at different initial concentrations (cr0) in the reaction tank

r∆t 0.0011 0.0057 0.0113 0.0225 0.0335 0.0445 0.0553

RI PT

r∆t 0.0009 0.0047 0.0093 0.0185 0.0276 0.0367 0.0456

cr0 = 200 µg L-1 rout =rin ∆c (µg L-1) 0.0004 0.09 0.0024 0.48 0.0047 0.94 0.0093 1.86 0.0138 2.76 0.0184 3.68 0.0230 4.60

SC

r∆t 0.0006 0.0030 0.0059 0.0118 0.0176 0.0234 0.0292

cr0 = 300 µg L-1 rout =rin ∆c (µg L-1) 0.0003 0.09 0.0014 0.45 0.0030 0.89 0.0059 1.77 0.0088 2.64 0.0118 3.53 0.0147 4.41

M AN U

Time step (min) 0.1 0.5 1 2 3 4 5

cr0 = 100 µg L-1 rout =rin ∆c (µg L-1) 0.0006 0.06 0.0028 0.28 0.0057 0.57 0.0113 1.13 0.0168 1.68 0.0223 2.23 0.0278 2.78

EP

TE D

Note: ∆c stands for the difference between the concentrations of naphthalene in seawater that remains in the reaction tank and in seawater that flows out of the reaction tank. It can be calculated with ∆c = cr0*( r∆t - rout).

AC C

611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627

29

ACCEPTED MANUSCRIPT

Table 3. Statistical results of the Monte Carlo Simulation Flow rate (L min-1) UV lamp Cost ($) Min 0.100 5 11.449 Max 0.495 6 14.000 Mean 0.315 5.977 12.797 Std 0.115 0.153 0.728 95% percentile [0.114, 0.478] [6, 6] [11.538, 13.823]

Table 5. Comparison between hourly stage and two stage optimization Time (hr) Optimization results ($) Hourly stage Two stage 32 8.9870 9.0407 30 8.6753 9.4853 28 8.7885 9.0622 27 8.6280 9.2705 26 9.0315 8.9188 25 9.0442 9.0442

636 637

AC C

EP

634 635

Table 4. Sensitivity analysis of the number of optimization stages with 36 hours The number Stage Optimization of stages duration (hr) results ($) 36 1 9.1125 18 2 9.2503 12 3 8.9999 6 6 9.1148 2 18 8.3622 1 36 11.4480

TE D

632 633

M AN U

SC

630 631

Table 2. The comparison of minimized treatment cost based on different Np and Ng combinations Test # Np Ng Min Cost ($) 1 30 50 9.6784 2 30 100 9.1125 3 30 200 9.1125 4 50 100 9.3094 5 100 100 9.5404

RI PT

628 629

30

Fig. 1. Flow chart of the simulation-based dynamic mixed integer nonlinear programming (SDMINP) approach

EP

639 640

AC C

638

TE D

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

31

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

641

TE D EP

643

Fig. 2. A schematic plot of the flow-through UV treatment system of the case study

AC C

642

32

EP

Fig. 3. Experimental and ANN modeling results at (a) 10 and 500 µg L-1 initial concentrations with 40 oC, 32.5 ppt and 8.27 mW cm-2; (b) 25, 32.5, and 40 ppt salinity with 20 oC, 10 µg L-1 and 2.88 mW cm-2; (c) 2.88 and 8.27 mW cm-2 fluence rate with 20 oC, 25 ppt and 10 µg L-1; and (d) 23 and 40 oC temperature with 40 ppt, 10 µg L-1 and 8.27 mW cm-2

AC C

644 645 646 647 648 649

TE D

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

33

EP

TE D

Fig. 4. Variations of the concentrations in the storage and reaction tanks and the treatment cost during the 36-hour UV exposure

AC C

650 651 652 653

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

34

EP

TE D

Fig. 5. Variations of the flow rate and the number of UV lamps during the 36-hour UV exposure

AC C

654 655 656 657

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

35

EP

TE D

Fig. 6. Variations of the concentrations in the storage and reaction tanks and the treatment cost during the 36-hour UV exposure using the single-stage optimization

AC C

658 659 660 661

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

36

EP

TE D

Fig. 7. Minimization of treatment time and cost using both multi-stage (SDMINP) and singlestage optimization

AC C

662 663 664

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

37

ACCEPTED MANUSCRIPT

EP

TE D

M AN U

SC

RI PT

Process simulation and control of marine oily wastewater treatment using UV; Simulation-based dynamic mixed integer nonlinear programming (SDMINP); Integrated process simulation, genetic algorithm and multi-stage programming; SDMINP performs better than the traditional static single-stage optimization.

AC C

• • • •