chemical engineering research and design 8 9 ( 2 0 1 1 ) 665675

Contents lists available at ScienceDirect

Chemical Engineering Research and Design

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

Taguchis parametric design approach for the selection of optimization variables in a refrigerated gas plant
Nooryusmiza Yusoff , M. Ramasamy, Suzana Yusup
Department of Chemical Engineering, Universiti Teknologi Petronas, Bandar Seri Iskandar, 31750 Tronoh, Perak, Malaysia

a b s t r a c t
This paper presents a systematic procedure for selecting optimization variables in a refrigerated gas plant (RGP) using the Taguchi method. A dynamic RGP model developed under HYSYS environment is utilized as a test bed. Nine variables with three levels each are employed for optimizing RGP prot. These optimization variables are selected due to their roles as manipulated variables in controlling the process. Results are validated by comparing optimum values of RGP prot obtained through a set of HYSYS experiments and those from analyses of means. Crown Copyright 2010 Published by Elsevier B.V. on behalf of The Institution of Chemical Engineers. All rights reserved. Keywords: Taguchi method; Parametric design; Gas plant; Optimization

1.

Introduction

Natural gas has emerged as an important source of cleaner energy. Improving operational efciency of a refrigerated gas plant (RGP) may signicantly increase its prot margin. This proves to be a challenging task due to the time-varying nature of feedstock ow rates and compositions. At the plant outlet, product values may also uctuate at frequent intervals depending on market conditions. Such challenges in operation and economics are typically handled by employing advanced process control (APC) and real-time optimization (RTO), respectively. A systematic approach for selecting control and optimization variables is required for the effective implementations of APC and RTO. One of the most widely employed methods is called Taguchi design of experiments. Parametric design through the Taguchi method was successfully applied in many engineering disciplines. For example, Cheng et al. (2008) studied thermal chemical vapor decomposition of silicon lm by integrating computational uid dynamic codes in FLUENT and a dynamic model of Taguchi method with L18 (21 37 ) orthogonal arrays. They found that thickness deviation of silicon lm could be reduced

by up to 11% points from 36% previously. Engin et al. (2008) employed L16 (42 22 ) arrays to investigate color removal from textile dyebath efuents in a zeolite xed-bed reactor. Hexadecyl-trimethyl-ammonium bromide (HTAB, C19 H42 BrN), as a surfactant, was added to increase the absorption capacity of the zeolite. Experimental results indicated that HTAB concentration, zeolite bed height and wastewater owrate were important parameters, whereas HTAB owrate was an insignicant parameter. In another application, Chiang (2005) studied the cooling performance of a parallel-plain n heat sink module using L18 (21 37 ) arrays. Through the analysis of variance (ANOVA), four out of eight variables were found to exhibit signicant contribution to the cooling process. The signicant variables are number of opening slots (34.8%), surface area of copper base (22.7%), fan capacity (13.6%) and height of n ake (8.7%). By utilizing an optimal conguration of these variables in a simulated environment, a 15% improvement in the cooling performance was achieved. Lee and Kim (2000) proposed a controller gain tuning technique for multi-axis PID control system. The test bed was a parallel-mechanism machine tool containing eight servodrivers. Each servodriver

Abbreviations: ANOM, analysis of means; ANOVA, analysis of variance; APC, advanced process control; DOF, degrees of freedom; FC, ow controller; FG, feed gas; MSD, mean squared deviation; PC, pressure controller; PFD, process ow diagram; PID, proportional-integralderivative; PRU, product recovery unit; RC, ratio controller; RGP, refrigerated gas plant; RTO, real-time optimization; SC, surge controller; SNR, signal-to-noise ratio; SRC, split-range controller; TC, temperature controller. Corresponding author. Tel.: +60 5368 7584; fax: +60 5365 6176. E-mail addresses: nooryus@petronas.com.my, nooryus@gmail.com (N. Yusoff). Received 4 July 2009; Received in revised form 28 September 2010; Accepted 30 September 2010
0263-8762/$ see front matter Crown Copyright 2010 Published by Elsevier B.V. on behalf of The Institution of Chemical Engineers. All rights reserved.

doi:10.1016/j.cherd.2010.09.021

666

chemical engineering research and design 8 9 ( 2 0 1 1 ) 665675

Table 1 Feed gas (FG) compositions. Nomenclature Cm k E Em K L M N NR P R Rk RM Sk

m Vk m x m xkl

Component
Methane Ethane Propane i-Butane n-Butane i-Pentane n-Pentane n-Hexane Nitrogen Carbon dioxide

FG-A
0.8956 0.0525 0.0289 0.0103 0.0060 0.0003 0.0002 0.0001 0.0040 0.0020

FG-B
0.8790 0.0516 0.0284 0.0101 0.0059 0.0003 0.0002 0.0001 0.0039 0.0206

FG-C
0.8465 0.0497 0.0273 0.0098 0.0057 0.0003 0.0002 0.0001 0.0037 0.0567

FG-D
0.7604 0.1581 0.0441 0.0080 0.0051 0.0000 0.0000 0.0000 0.0177 0.0066

m xk
m xopt

percentage contribution of factor k in Case m (%) expenses (RM/min) error for Case m (%) number of factors number of levels number of experiments in external array number of experiments in the internal array number of repeated level prot (RM/min) revenues (RM/min) ranking of factor k Ringgit Malaysia sum of squares of prot due to factor k [(RM/min)2 ] variance of factor k in Case m [(RM/min)2 ] average of prot in Case m (RM/min) average of prot due to factor k at level l in Case m (RM/min) average of prot due to factor k over all levels L in each Case m (RM/min) optimal prot in Case m (RM/min)

2.

RGP process and control descriptions

Indices Adj k l m n

adjusted factor level column in external array or case number row in internal array or experiment number

has four controller gains, thus a total of 32 gains needed to be tuned simultaneously. By utilizing an L9 (34 ) orthogonal array, robust controller gains were obtained. A performance indicator namely the index of average position and velocity errors was reduced by 61.4%. In addition, average signal-to-noise ratio (SNR) was increased by 8.5 dB to attain better control of the machine tool. In the current work, the Taguchi method for the design of experiment, signal-to-noise ratio (SNR), analysis of means (ANOM) and analysis of variance (ANOVA) is applied to accomplish the following objectives (Roy, 1990):

1. To identify and rate optimization variables under the inuence of disturbance variables. 2. To determine optimum congurations of optimization and disturbance variables. 3. To estimate and validate the maximum RGP prot within the specic constraints of all variables.

Cross-array experiments employing L27 (37 ) and L9 (32 ) internal and external arrays, respectively, are performed on the RGP dynamic model to investigate the effects of seven optimization variables and two disturbance variables at 3 levels each on the RGP prot. These variables are found to strongly inuence the RGP prot in a co-current study. The variables and RGP process are described in the next section.

A refrigerated gas plant (RGP) processes raw feed gas into sales gas and natural gas liquids. Main feeds to the RGP come from three streams, namely, A, B and C. Another feed, stream D, is available in smaller quantities to boost the gross heating value of the sales gas. Feed gas compositions are presented in Table 1. Water, sulfur and mercury are assumed to be absent when the feed gas enters the RGP. In other words, all feed gas streams are assumed sweet and dry with varying levels of carbon dioxide content. At base-case operation, mixed feed gas ows at 250 ton per hour (ton/h), and enters the RGP at 20 C and 50 bar. It is cooled by exchanging heat with the processed gas in three cold boxes (E-101, E-103 and E-105), a propanerefrigeration cooler (E-102) and an air cooler (E-106). The feed gas is ashed in two stages in order to enhance vaporliquid separation. The resulting vapor is expanded in a turboexpander (KT-101) and/or JouleThompson (JT) valve to improve natural gas liquids recovery. Liquids are fed to different stages of demethanizer (C-101). Bottom product of the demethanizer is sent to Product Recovery Unit (PRU) for further processing of the natural gas liquids. The top product of the demethanizer is sent to an absorber (C-102) to enhance recoveries of ethane and heavier hydrocarbons. The top product of the absorber is now called sales gas and leaves the column at about 20 bar. Recompression of the sales gas is carried out twice at compressors K-101 and K-102 to meet the pipeline pressure specication of above 30 bar. The bottom product of the absorber containing an equal amount of methane and natural gas liquids is recycled to the top of the demethanizer for further processing. A dynamic model of the RGP under HYSYS environment has been developed by Yusoff et al. (2008). The process ow diagram (PFD) of this model is shown in Fig. 1. This large model contains 762 variables and 21 regulatory control loops. The controllers and the energy streams are numbered systematically. However, numbering for most material streams is omitted for the clarity of PFD reading. Key controllers are: (a) a split-range controller SRC103 to control the feed gas stream temperature after cold box E-101, (b) a temperature controller TC101 to control the cooler E-102 outlet temperature, (c) a split-range controller SRC102 to control the feed gas stream temperature after the cold box E-103, (d) a ratio controller RC101 to control the ow rate of the gas to the absorber C-102 over that to the turboexpander KT-101, (e) a ow controller FC104 to control the ow of rich feed gas stream (FG-D) from an adjacent plant used mainly to increase the sales gas gross heating values, (f) a pressure controller PC101 to control the demethanizer C-101 overhead pressure, (g) a temperature controller TC102 that manipulates the demethanizer reboiler

chemical engineering research and design 8 9 ( 2 0 1 1 ) 665675

667

Fig. 1 RGP process ow diagram. capacity, and (h) a ow controller FC101 to control the plant load. The major feed gas streams are FG-A, FG-B and FG-C comprising different levels of carbon dioxide contents. Two product streams, namely SG and NGLs, contain sales gas and natural gas liquids, respectively. Important energy streams are described in the next section. Since hundreds of variables are available for manipulation, proper selection of optimization variables is required. This task may be performed using the Taguchi method. called system design, deals with working levels of design factors. Here, the development and testing of a system are performed based on the scientic and engineering knowledge given the limitations of the current technology. The second stage is termed parametric design. At this stage, an optimum condition is determined at specic factor levels with or without the presence of uncontrolled factors. The third stage is tolerance design, which is used to ne tune the optimum factor levels obtained from the parametric design stage. This is important to tighten the product quality, reduce both the capital and the operating costs, and increase customer satisfaction index at the same time. The terms factors, variables and parameters are synonymously used throughout this paper. They are the causes that affect the outcomes of certain processes. These factors can be divided into controllable factors and noise factors. Both controllable and noise factors are specied and maintained at specic levels during the experimental stage. The difference between the two factors is that the former are employed to manipulate the process whereas the latter are used to assess the degree of severity of their inuence to the process. Due to this reason, two orthogonal arrays namely the internal and external arrays are used for the controllable and noise factors, respectively. This approach, termed cross-array experiments, can determine the optimum design conditions with minimal inuence of the noise factors to the process outcomes.

3.
3.1.

Taguchi method
An overview

Prior to the emergence of Taguchi methods, factorial design of experiments was employed to investigate all the possible combinations of parameters that can fabricate the best product at certain costs. When the number of parameters is large, the factorial design approach faces four major drawbacks namely (Roy, 1990): (1) It would be costly and time consuming to perform all the experiments; (2) Completely different results may be obtained from two designs of the same experiment; (3) Contribution of each parameter cannot be determined; and (4) Results may be difcult to be interpreted. The Taguchi methods were developed to overcome these limitations by providing a systematic approach to the design of experiments. Standard tables known as the orthogonal arrays are used to ease the design of experiments and to obtain consistent results from the same experiment. An orthogonal array is abbreviated as LN , where the subscript N refers to the number of trials that needs to be run for a given experiment. To be more descriptive, the number of levels and factors are included in a parenthesis next to the abbreviation of the orthogonal array. For example, L4 (23 ) Taguchi method refers to using a four-trial orthogonal array to investigate the inuence of three factors at two levels each on a given process. The Taguchi methods serve as an off-line tool for designing quality into products in a three-stage process. The rst stage,

3.2.

A case study

The implementation of the Taguchi method can be illustrated with the help of a ow diagram shown in Fig. 2. The procedure employed in the current work is modied from that presented by Yang et al. (2007). It is more compact and includes a failure loop for invalidated design of experiment. Step 1 is the problem formulation, which requires dening an objective function, factors and levels. This step requires an in-depth knowledge of the process of interest. In the case of RGP, inputs from experienced operators are essential in determining the

668

chemical engineering research and design 8 9 ( 2 0 1 1 ) 665675

Start Objective function Factors Levels Orthogonal array Conduct experiment SNR ANOVA Response plot

Table 3 Economics data. Component

FG-A FG-B FG-C FG-D Sales gas Ethane Propane Butane Condensates Refrigeration Steam duty Compressor duty Turboexpander-compressor duty Electricity

Price
6.0 5.5 5.0 7.0 14.4 304.0 569.0 908.0 673.8 169.0 89.9 84.5 42.3 233.3

Unit
RM/MMBtu RM/MMBtu RM/MMBtu RM/MMBtu RM/MMBtu RM/ton RM/ton RM/ton RM/ton RM/MWh RM/MWh RM/MWh RM/MWh RM/MWh

Problem Formulation

Experimental Design No

Analysis of Results

Validated? Yes End

Fig. 2 Flow diagram of Taguchi method. potential optimization and disturbance variables, as well as their normal, high and low values (Table 2). The optimization and disturbance variables are, respectively, called controllable and noise factors in the Taguchi-related literature. Limits of high and low values are called levels. Median, and lower and upper quartiles of the limits may also be included to augment experimental design congurations. Controllable factors A to G are set up in an L27 (37 ) internal array while noise factors H and I in an L9 (32 ) external array. Cross-array experiments between the external and internal arrays are conducted to include the effect of noise on the controllable factors (Taguchi and Konishi, 1987). Outputs from running these experiments are prot values obtained from an objective function (Eq. (1)).
I J

P=
i=1

Ri
j=1

Ej

(1)

where P is the prot, Ri (I = 5) are the revenues and Ej (J = 10) are the expenses. In other words, the objective function is specially formulated as an economic expression of product values and operational expenses. Valuable products are the sales gas and the natural gas liquids comprising ethane, propane, butane and condensates. Revenues are calculated based on the ow rates of the respective products. The operational expenses are mainly due to the costs of feed gas and utilities in the forms of refrigeration cooler duty (E-102Q), demethanizer reboiler duty (E-104Q), compressor fuel gas consumption (K102Q), turboexpander-compressor maintenance (KT/K-101Q) and electricity usage for pumping actions (P-101Q and P-102Q). The prices and corresponding units of each component of revenues and expenses are shown in Table 3. Since the RGP dynamic model is developed based on the rst principles, sig-

nicance of optimization and disturbance variables can be systematically established through the Taguchi method. Step 2 involves designing and conducting experiments. The number of factors and levels has an effect on selection of standard orthogonal arrays. For 7 controllable factors at 3 levels such as the one in this study, an L27 array consisting of 27 rows and 13 columns is appropriate. The rows and columns represent experimental runs and factors, respectively. Since only 7 controllable factors are used to calculate the RGP prot, the remaining 6 columns on the far right of the array are ignored. Similarly, an L9 array consisting of 9 rows and 4 columns is selected for 2 noise factors at 3 levels. Only the rst 2 columns are used in this study. Reduced arrays and outputs are shown in Tables 4 and 5. It should be noted that, in order to capture adequate responses of all 7 controllable factors towards the objective function, only 27 experiments from the internal array need to be conducted for each run from the external array. Hence a total of 243 (=27 9) experiments need to be performed. This is more appealing than running 19,683 (39 ) experiments under the full factorial design approach. Step 3 deals with the analysis of results. Three major statistical tools commonly applied in the Taguchi method are signal-to-noise ratio (SNR), analysis of means (ANOM) and analysis of variance (ANOVA). SNR for each Run n is dened as:

(SNR) = 10 log (MSD) ,

n = 1, . . . , N

(2)

where N = 27 is the number of experiments in the internal array. Since the RGP prot is chosen as the objective function

Table 2 Description of factors and levels for RGP. Factors

A B C D E F G H I

Level 1
90 30 90 0.005 0 22 70 250 6.0

Level 2
95 40 95 0.075 25 23 85 280 5.5

Level 3
100 50 100 0.15 50 24 100 310 5.0

Units
% % % % barg % ton/h RM/MMBtu

Description
Split range controller SRC103 output Temperature controller TC101 output Split range controller SRC102 output Ratio controller RC101 ratio Flow controller FC104 output Demethanizer C-101 overhead pressure Temperature controller TC102 output Feed gas ow rate Feed gas prices

chemical engineering research and design 8 9 ( 2 0 1 1 ) 665675

669

Table 4 Taguchi internal arrays showing levels of controllable factors. Run A

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3

Factors B
1 1 1 2 2 2 3 3 3 1 1 1 2 2 2 3 3 3 1 1 1 2 2 2 3 3 3

C
1 1 1 2 2 2 3 3 3 2 2 2 3 3 3 1 1 1 3 3 3 1 1 1 2 2 2

D
1 1 1 2 2 2 3 3 3 3 3 3 1 1 1 2 2 2 2 2 2 3 3 3 1 1 1

E
1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3

F
1 2 3 1 2 3 1 2 3 2 3 1 2 3 1 2 3 1 3 1 2 3 1 2 3 1 2

G
1 2 3 1 2 3 1 2 3 3 1 2 3 1 2 3 1 2 2 3 1 2 3 1 2 3 1

Eq. (4), n is an index referring to the run numbers in L27 array that correspond to unique congurations of factorial levels. This means that n = 1,. . .,NR may not follow a sequence from 1 to NR . For example, index n = 1 may refer to the tenth row in L27 array containing the rst output of a particular factor at level 2. After determining the average of factor k at level l, m the average of factor k over all levels L in each Case m, xk is calculated as: m xk = 1 L
L

m xkl ,
l=1

k = 1, . . . , K

(5)

m These two averages are used to calculate the variance Vk which has two contributors. The numerator is the sum of the m m squares between the two averages of factor k (xkl and xk ). The denominator is called the degrees of freedom of factor k over m all levels L in Case m, (DOF)k . L 2 m (xm xk ) l=1 kl , m1 Lk

m Vk =

k = 1, . . . , K

(6)

The percentage contribution Cm is obtained by dividing the k m individual variance of factor k, Vk from the total variance of all factors and multiplying the result with 100. Cm = k
m 100Vk K Vm k=1 k

k = 1, . . . , K

(7)

in this work, the mean squared deviation (MSD) is dened to uphold the-larger-the-better quality principle as follows: 1 M
M

(MSD) =

1 (xmn xm )
2

n = 1, . . . , N

(3)

m=1

where M = 9 is the number of experiments in the external array. Eq. (3) is slightly modied from the one presented by Roy (1990). This is necessary to avoid dealing with large absolute values of the RGP prot. In the current work, deviational values of prot are obtained from the spread around means xm of respective cases. For the ANOM and ANOVA, two averages must be calculated m a priori. Average of factor k at level l in Case m, xkl , is taken as the sum of respective factorial values divided by the number of repeated levels, NR . 1 NR
NR mn xkl , n=1

Step 4 is the validation of experiment. For each run in the external array, there are 27 experimental runs in the internal array. By design, only one run will yield the highest prot margin. Preliminary visual inspection of trends of each factor average contributions at all levels can be made through a response plot. Here, average values of prot and SNR of factor k at levels l = 1,. . .,L (L = 3) are plotted against the corresponding factors. The response plot may be used to locate optimal design conguration for the purpose of verifying results. Additional experiments with 9 runs are required to compare both the experimental and the calculated values of prot. The calm culated optimum prot xopt in Case m is obtained by summing m for the same case with maximum devi up the global mean x m ations of average values of factor k at level l, xkl from the m. corresponding average values at all levels, xk
K m xopt = xm + k=1

m m max(xkl ) xk

l = 1, . . . , L

(8)

where k = 1, . . . , K; l = 1, . . . , L; m xk = xm , (4) m Eq. (9) implies that the average of factor k over all levels, xk m of the respective Case m. This equals to the global mean x is true since the Taguchi design of experiment is unbiased by construction. k = 1, . . . , K (9)

m xkl =

m = 1, . . . , M

where NR = 9, K = 7 and L = 3 are correspondingly the number of repeated levels, factors and levels. It should be noted that in

Table 5 Taguchi external arrays showing levels of noise factors. Factor 1

H I 1 1

4.

Results and discussion

Run 2
1 2

3
1 3

4
2 1

5
2 2

6
2 3

7
3 1

8
3 2

9
3 3

All experiments are conducted using the RGP dynamic model developed under HYSYS environment. The prot (Eq. (1)) is calculated online in a built-in spreadsheet. The material and energy units are converted a priori to be consistent with the basis of 1-min interval calculation. To ensure repeatability,

670

chemical engineering research and design 8 9 ( 2 0 1 1 ) 665675

Table 6 Results of Taguchi crossed-orthogonal-array experiments. Run x1n

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 Mean 1742.94 1764.03 1798.38 1731.30 1754.64 1777.01 1723.50 1747.93 1763.72 1760.99 1744.46 1759.31 1738.38 1728.07 1742.37 1732.66 1720.47 1733.46 1730.56 1739.71 1728.68 1723.58 1733.20 1720.77 1706.99 1717.47 1705.45 1739.63

Prots (RM/min) x2n

1785.86 1805.35 1832.00 1778.85 1800.10 1817.79 1773.30 1794.78 1808.65 1802.53 1791.71 1805.31 1786.61 1778.10 1791.84 1782.34 1770.57 1785.18 1779.28 1789.44 1779.92 1774.15 1784.63 1770.29 1760.48 1773.86 1758.57 1787.46

x3n
1741.67 1755.75 1777.96 1735.15 1751.35 1760.46 1729.05 1748.01 1757.94 1746.53 1743.56 1758.48 1740.90 1734.41 1749.59 1736.92 1728.48 1743.70 1737.60 1745.09 1737.27 1731.25 1741.88 1732.80 1722.91 1735.39 1723.00 1742.48

x4n
1973.24 1996.33 2019.41 1957.65 1982.67 2002.25 1950.23 1974.82 1996.29 1982.39 1969.70 1982.90 1969.89 1951.93 1968.70 1960.44 1944.59 1959.46 1952.84 1965.75 1948.20 1948.44 1959.36 1942.57 1930.70 1942.55 1929.39 1965.28

x5n
2007.01 2026.64 2038.59 1991.79 2015.66 2030.58 1985.75 2009.52 2027.19 2013.95 2005.60 2019.01 2001.48 1988.02 2003.45 1996.18 1981.31 1996.17 1992.14 2001.44 1986.54 1987.44 1997.94 1981.62 1970.28 1983.44 1969.82 2000.32

x6n
1947.93 1965.99 1977.41 1942.41 1959.77 1973.66 1936.35 1955.43 1970.67 1958.16 1951.41 1965.31 1948.91 1939.62 1955.76 1944.51 1934.02 1950.31 1941.69 1953.61 1941.44 1938.21 1949.52 1937.03 1927.69 1940.31 1928.74 1949.48

x7n
2189.85 2213.38 2230.85 2177.60 2201.50 2222.27 2161.30 2186.93 2209.06 2199.29 2179.52 2199.07 2186.01 2161.86 2186.88 2175.24 2155.48 2175.49 2162.20 2181.14 2157.78 2156.62 2167.50 2153.45 2139.05 2154.70 2138.05 2178.59

x8n
2225.79 2249.27 2260.46 2212.26 2238.52 2253.86 2198.99 2224.65 2243.64 2232.72 2218.32 2237.94 2222.60 2201.19 2225.69 2212.85 2195.28 2213.85 2203.57 2221.69 2199.03 2199.82 2210.99 2195.00 2181.13 2196.45 2181.83 2216.94

x9n
2160.85 2179.83 2192.79 2150.55 2169.87 2185.60 2143.29 2163.69 2180.47 2166.66 2160.46 2173.00 2159.27 2148.40 2166.81 2151.70 2142.96 2158.20 2150.47 2160.53 2149.73 2147.33 2155.37 2145.96 2130.81 2146.72 2136.44 2158.44

xn
1975.01 1995.18 2014.21 1964.17 1986.01 2002.61 1955.75 1978.42 1995.29 1984.80 1973.86 1988.92 1972.67 1959.07 1976.79 1965.87 1952.57 1968.42 1961.15 1973.16 1958.73 1956.32 1966.71 1953.28 1941.12 1954.54 1941.25 1970.96

Note: Superscript n refers to experimental run number in internal arrays.

HYSYS 2006 SP5 running on Windows XP Professional operating system is employed. This process simulator is installed in a standard Dell Optiplex GX520. Reproducibility is also ensured because the experimental steps are pre-congured in an Event Scheduler. The changes on levels of all factors are set to run in parallel. The experiments are stopped after all the factors reach steady-state at 420 min simulation time. In most experiments, the prot values level off after 360 min but in some runs, the values slightly uctuate towards the end. For these runs, the last 60 values are averaged out. Therefore, only one prot value is obtained for each run. Major results are presented in Table 6. For convenience, the prot is denoted as xmn where m (m = 1,. . .,9) and n (n = 1,. . .,27) are indices of external and internal runs, respectively. Cases 19 refer to the corresponding values of index m. In the following discussion, the Cases 13, 46 and 79 are categorized as Groups I, II and III, respectively, due to the similarity in conguration of noise factor H (plant load) and sequence of noise factor I (feed gas prices).

is caused by the presence of noise factor H, which is the plant load. Increasing the plant load increases the amount of feed gas and thus RGP prot due to additional production of sales gas and natural gas liquids. However, it should be noted that the RGP was designed to process a maximum of 310 ton/h of feed gas. Any amounts higher than this value will push the equipment loads towards the upper constraints. On the other hand, the RGP load can be reduced to 100 ton/h of feed gas without the need for total plant shutdown. However, underloading is undesirable since the RGP prot will also diminish. Similarly, the effect of noise factor I on the RGP prot can be deduced. The highest values of average prot in each Group I, II and III are generated from I2 (factor I, level 2) conguration. As clearly shown in Cases 2, 5 and 7, the average prots are, respectively, 1787.46, 2000.32 and 2178.59 RM/min. This is attributed to the different economic values of the feed gas streams FG-A, FG-B and FG-C due to different values of carbon dioxide contents as presented previously. Highly priced stream FG-A decreases the RGP prot while the cheaper stream FG-C increases it.

4.1.

Effect of noise factors 4.2. Averaged prot analysis

Global means for Cases 1, 2 and 3 (Group I) are, respectively, 1739.63, 1787.46 and 1742.48 in the unit of Ringgit Malaysia (RM)/min. For Cases 4, 5 and 6 (Group II), the global means are, respectively, 1965.28, 2000.32 and 1949.48 RM/min. In the last three cases (Group III), the values are 2178.59, 2216.94 and 2158.44 RM/min. The means for Groups I, II and III can be calculated as 1756.53, 1971.69 and 2184.66 RM/min, respectively. A difference of about 200 RM/min is noticed between Groups I and II as well as between Groups II and III. This discrepancy

To study relative signicance of factors quantitatively, ranking of factors is initially performed using analysis of mean (ANOM) for the RGP prot. Eqs. (4) and (5) are applied to calculate the averages of factor k. For the RGP prot analysis, the values of prot from Cases 1 to 9 are averaged out in a row-by-row basis as implemented by Yang et al. (2007). This approach is consistent with the general procedures of the Taguchi method that deals with average values. Superscript m referring to run

chemical engineering research and design 8 9 ( 2 0 1 1 ) 665675

671

Table 7 Analysis of means for average prot. xAl

Level 1 Level 2 Level 3 xk Dk Ek Rk 1958.85 1949.78 1939.80 1949.48 9.37 19.04 1

xBl
1955.88 1949.43 1943.11 1949.48 6.41 12.77 3

xCl
1949.44 1949.72 1949.28 1949.48 0.24 0.44 7

xDl
1948.04 1949.05 1951.34 1949.48 1.87 3.30 5

xEl
1942.87 1949.96 1955.59 1949.48 6.12 12.72 4

xFl
1949.06 1948.89 1950.49 1949.48 1.01 1.60 6

xGl
1939.88 1951.13 1957.42 1949.48 7.94 17.54 2

number in the external array may be omitted in this section. Hence, the average values of prot are denoted as xn . Ranking is initially determined from ANOM, which shows the deviation of the highest value from the lowest value of average of factor k at level l: l = 1,. . .,L (L = 3). This deviation is denoted as Ek . The highest ranking is assigned to a factor carrying the highest Ek value. In the case of 7 controllable factors that affect the RGP prot, the descending order of importance is AGBEDFC (Table 7). This implies that factor A is the most signicant while factor C is the least signicant. Ranking from ANOM is veried with ranking from analysis of variance (ANOVA). In the latter analysis, the averages of factor k at level l and at all levels L are used to determine the sum of squares of factor k. In general, an Sk value represents spread of experimental results of factor k from the data average. A large value such as the one obtained here indicates the signicant contribution of that particular factor towards the output (RGP prot). On the other hand, a factor is deemed unimportant if its Sk value approaches zero. Since Sk values are unbounded at the higher end, it is convenient to denote relative importance of a factor k in terms of its percentage contribution, Ck (Eq. (7)). Before calculating Ck , a quantity called general variance of factor k, Vk (Eq. (6)) needs to be determined. This quantity differs from the population variance 2 , which could only be obtained if all 19,683 (39 ) possible experiments had been conducted. Vk is obtained from Eq. (6) where the degree of freedom of factor k (DOF)k is one less its number of levels. Ranking of factor k, Rk from ANOVA is presented in Table 8. The descending order of importance of 7 controllable factors is AGBEDFC. This result is identical to the one obtained from ANOM. Factor A (SRC103 output) is a major contributor with 35.6%. This factor is employed to control the temperature of the feed gas exiting cold box E-101 that could inuence refrigeration cooler duty (TC101 output). Lower feed gas temperature means less cooler E-102 duty is required to maintain the same separation and thus higher RGP prot. Contribution of factor B (TC101 output) at 16.0% is about half of that for factor G (TC102 output) at 31.0%. Both the factors B and G play an important role in the plant energy balance. Hence, an increase in TC101 output that further reduces the RGP temperature must be accompanied by a similar rise in TC102 output. Disproportionate contribution of factors B and G are attributed to the presence of factor A.

The effect of factor E (FC104 output) is also signicant with 16.0% contribution. Flow controller FC104 output controls reinjection of rich hydrocarbon to boost sales gas gross heating value. An increase in FC104 output increases the sales gas production. Due to the large difference of prices between the sales gas and FG-D, an increase in the sales gas production rate increases the RGP prot. However, it should be noted that only 10 ton/h of FG-D is available for this purpose. Contribution of factor F (pressure controller PC101 that controls demethanizer overhead pressure) is fteen times as large as that of factor C (SRC102 output). Both the factors C and F are mainly used to control the demethanizer overhead quality. Since the installation of a section called gas subcooled process in the RGP, inuence of split-range controller SRC102 controller on the demethanizer overhead quality has been drastically reduced and taken over by the pressure controller PC101. Contribution of factor D (RC101 controller ratio) is minor at 1.1%. This factor controls split of the processed gas going into the gas subcooled process section and/or turbo-expander. A higher value of RC101 controller ratio promotes more recovery of ethane and heavier hydrocarbons at the gas subcooled process section. In other words, factor D is useful in improving ethane recovery but only slightly signicant in increasing the RGP prot due to the lower price of ethane as compared with the other natural gas liquids values.

4.3.

Signal-to-noise ratio (SNR) analysis

The RGP prot is determined by independently setting 7 controllable and 2 noise factors in the Taguchi arrays. An L27 (37 ) internal array is selected for controllable factors and an L9 (32 ) external array for noise factors. This means that each experimental run from the internal array need to be repeated 9 times under different congurations of the external array. In the previous section, signicance of controllable factors in maximizing the RGP prot has been discussed with moderate inuence of noise factors. In this section, signicance of controllable factors is investigated using signal-to-noise ratio (SNR) approach. Qualitative principle in the form of thelarger-the-better SNR is selected for this analysis. A high value of SNR in a particular run indicates the minimum effects of noise factors on the outputs. The SNR values for each run are calculated using Eqs. (2) and (3), and the results are pre-

Table 8 Analysis of variance for average prot. A

Sk (DOF)k Vk Ck Rk 4898.45 2 2449.23 35.61 1

B
2201.90 2 1100.95 16.01 3

C
2.69 2 1.34 0.02 7

D
154.57 2 77.28 1.12 5

E
2193.95 2 1096.97 15.95 4

F
41.77 2 20.89 0.30 6

G
4261.94 2 2130.97 30.98 2

672

chemical engineering research and design 8 9 ( 2 0 1 1 ) 665675

Table 9 Results of signal-to-noise ratio (SNR) analysis. Run

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27

(SNR)n
5.32 26.40 32.03 8.89 22.37 29.11 23.51 16.88 27.15 19.09 4.43 24.91 0.58 20.83 12.91 13.49 24.95 3.51 18.34 12.84 19.31 22.69 18.06 23.72 28.86 22.06 28.63

(SNR)Adj
23.37 44.46 50.09 26.95 40.43 47.17 41.57 34.94 45.21 37.15 22.49 42.96 18.64 38.89 30.97 31.54 43.00 14.54 36.39 5.22 37.37 40.74 0.00 41.78 46.91 40.11 46.68

sented in Table 9. Since an SNR value may be negative when MSD rises above 1, a bias is subtracted from all SNR values. The bias is chosen as the minimum value of SNR across all case studies. In essence, an adjusted SNR is dened by Eq. (10): (SNR)Adj = (SNR)
mn mn

min (SNR)

(10)

In the subsequent discussions, the word adjusted is omitted for convenience. All SNR values refer to adjusted SNR values. These values are used to determine ranking of controllable factors that can minimize effects of noise factors. Ranking is performed using a similar method employed in the averaged prot analysis. However, SNR values instead of averaged values of prot are used in carrying out ANOM and

ANOVA. For ANOM, averages of SNR of factor k at levels 1, 2 and 3 are initially calculated before determining the corresponding Ek values. Ranking of 7 controllable factors in descending order of importance is FEACBDG (Table 10). This shows that factors F and G are, respectively, the most and the least important factors. However, it is important to note that SNR results disagree with the ones based on the averaged prot. Interpretation of results from SNR analysis is deferred until after discussion on ANOVA, which is utilized to verify the results from ANOM. The main objective of ANOVA is to calculate the percentage contributions of factor k, Ck in inuencing the RGP prot. Ranking of factors may also be determined from Ck values. A factor with the highest Ck value is the most signicant. The procedure for calculating Ck values are presented previously in the case of averaged prot analysis and thus is not repeated here. The descending order of importance from ANOVA is FEACBGD (Table 11). The order is similar to that from ANOM except for the last two, in which factors D and G switch places. Since ANOVA is a second order statistical analysis, its results are more reliable than those of ANOM. It is evident that ranking orders for statistical analyses using the averaged prot and the SNR values differ markedly. Factor A (SRC103 output) with 35.6% contribution is the most signicant in maximizing RGP prot. On the other hand, factor F (PC101 set point) with 16.1% is the most important in minimizing effects of noise factors through SNR analysis. PC101 is a pressure controller responsible for maintaining the demethanizer overhead pressure. Raising PC101 set point from 22 to 24 bar increases the demethanizer top column temperature by about 5 C. This action induces losses of ethane and propane in the sales gas product stream. On the other hand, a low value of PC101 set point at 22 bar helps to improve the recovery of natural gas liquids due to cooler condition at the enriching part of the demethanizer. Given uctuating disturbances in the forms of noise factors H (plant load) and I (feed gas prices), the pressure controller PC101 should be set at appropriate level to stabilize the demethanizer. Factor E (FC104 output) at 14.8% contribution is the next important factor. Flow controller FC104 controls the rich hydrocarbon (FG-D) injection to the sales gas product stream. In the event of low plant load, FC104 output is increased to boost the sales gas ow rate and the gross heating values. This practice will drive revenue upwards due to the large gap between sales gas and FG-D

Table 10 Analysis of means for SNR. (SNR)Al

Level 1 Level 2 Level 3 xk Ek Rk 39.35 31.13 32.80 34.43 8.22 3

(SNR)Bl
33.28 31.73 38.28 34.43 6.55 5

(SNR)Cl
32.17 38.98 32.13 34.43 6.85 4

(SNR)Dl
37.79 31.40 34.09 34.43 6.39 6

(SNR)El
33.70 29.95 39.64 34.43 9.69 2

(SNR)Fl
25.08 37.00 41.21 34.43 16.13 1

(SNR)Gl
35.79 36.93 30.57 34.43 6.36 7

Table 11 Analysis of variance for SNR. A

Sk (DOF)k Vk Ck Rk 1019.78 2 509.89 11.67 3

B
633.06 2 316.53 7.24 5

C
840.08 2 420.04 9.61 4

D
555.96 2 277.98 6.36 7

E
1289.81 2 644.90 14.76 2

F
3781.63 2 1890.82 43.26 1

G
620.73 2 310.36 7.10 6

chemical engineering research and design 8 9 ( 2 0 1 1 ) 665675

673

prices. Factor A (SRC103 output) with 11.7% contribution is also signicant. Split-range controller SRC103 output controls the amount of processed gas entering and bypassing cold box E101. A high value of SRC103 output at 100% means no bypass. As a result, the hotter feed gas exchanges more heat with the much cooler processed gas and leaves cold box E-101 at lower temperature. This will have a positive effect on the recovery of natural gas liquids and thus the RGP prot. A stream with high carbon dioxide content contains lower quantity of hydrocarbons (termed leaner) to be processed. Since carbon dioxide has zero heating value, leaner feed gas requires less energy to be cooled and/or heated in order to achieve the same separation as the richer one. Thus SRC103 output value needs to be adjusted accordingly to handle specic feed gas stream FG-A, FG-B or FG-C. Explanation for signicance of factor C (SRC102 output) with 9.6% contribution is similar to that of factor A (SRC103 output). The only difference between the two factors is in regulating split of the sales gas ow at cold box E-103 instead of E-101. Contributions from factors B (TC101 output) and G (TC102 output) are about the same at 7.2% and 7.1%, respectively. The temperature controller TC101 output controls the propane refrigeration cooler duty while the temperature controller TC102 output controls the demethanizer reboiler duty. Both the variables control the amount of energy leaving or entering the RGP. A high value of plant load requires additional load on temperature controller TC101 duty to maintain the same recovery of natural gas liquids. Increasing the value of TC101 output must be matched by the same amount in TC102 output. With close to zero contribution, factor D (RC101 controller ratio) is the least signicant factor in regulating external disturbances. RC101 is a ratio controller that controls amount of feed gas entering turboexpander and gas subcooled process section of RGP. A high value of RC101 controller ratio improves the recovery of natural gas liquids. However, RC101 controller ratio can hardly be employed to minimize the effects of uctuations in the feed gas ow and the prices on to the RGP prot. Besides determining signicance of factors responsible for stabilizing disturbance, SNR results may also be employed to locate optimal conguration of factors to yield maximum RGP prot. This conguration is found based on the-higherthe-better qualitative principle. As shown previously, the highest SNR value of 50.1 comes from Run 3 with conguration A1 B1 C1 D1 E3 F3 G3 . This conguration consistently gives the highest prot values in all Cases 19 as shown by results in Run 3. In fact, a prot value of 2260.46 RM/min in Case 8 is the maximum one found from all the 243 experiments. On the other hand, a zero SNR value implies that the conguration A3 B2 C1 D3 E2 F1 G3 in Run 23 yields the lowest prot. However, it is found that an average prot value of 1966.71 RM/min at this run is closer to the global mean value of 1970.96 RM/min than to the minimum value of 1941.42 RM/min. This result conrms that the SNR principle employed in this study is only valid in determining the highest value but not the lowest value of the RGP prot. In general, the maximum and minimum values of prot can only be ascertained after running the entire 19,683 (39 ) experiments under the full factorial design approach. It is possible that prot values derived from the Taguchi method do not even reach either side of the extremities. For this reason, it is important to estimate and validate the maximum value of the prot by running another set of experiments under the optimal congurations of both controllable and noise factors.

1960

Profit (RM/min)

1955 1950 1945 1940 1935

A1 A2 A3 B1 B2 B3 C1 C2 C3 D1 D2 D3 E1 E2 E3 F1 F2 F3 G1 G2 G3

Average Factors at Respective Levels

Fig. 3 Response plot for averaged prot analysis.

4.4.

Validation

The nal step in the Taguchi method is a validation of results. The RGP prots obtained from conducting HYSYS experiments are compared with those calculated based on ANOM. In this case, ANOM results are obtained for individual Cases 19 as opposed to the one discussed previously where ANOM is performed using average values of the prot from all 9 cases. This is necessary because the latter lacks conguration of noise factors to be assigned in HYSYS Event Scheduler. ANOM prot values are obtained from Eqs. (4) to (8), which contains two terms, namely, the means and maximum differences of averages of factor k: k = 1,. . .,K (K = 7) for the corresponding Cases m = 1,. . .,9. Optimal congurations of controllable factors from analyses based on the averaged prot and the SNR values can be determined from two separate response plots (Figs. 3 and 4). Two qualitative observations can be made from this plot: (1) signicance of individual factors can be established from steepness of the factorial graph slopes and (2) levels of individual factors yielding the maximum RGP prots and/or the SNR can be visually determined. In the rst observation, the steeper the slope of a factorial graph, the more signicant the factor is. This deduction can complement percentage contributions of factors, Ck as discussed above. The second observation is important in estimating and validating the experimental results. A conguration of optimal levels of factors should yield a maximum value of the RGP prot. In the analysis that uses averaged values of prot, it is clear that the maximum prot could be obtained from the conguration A1 B1 C2 D3 E3 F3 G3 . This optimal conguration is set in HYSYS Event Scheduler and run for 420 min. The RGP prot is calculated automatically using built-in spreadsheet to prevent round-off error. The prot values of 1813.67, 1840.20, 1791.35, 2009.59, 2037.80, 1981.95, 2229.17, 2259.73 and 2192.06 RM/min
45

Signal-to-Noise Ratio

40

35

30

25

A1 A2 A3 B1 B2 B3 C1 C2 C3 D1 D2 D3 E1 E2 E3 F1 F2 F3 G1 G2 G3

Average Factors at Respective Levels

Fig. 4 Response plot for SNR analysis.

674

chemical engineering research and design 8 9 ( 2 0 1 1 ) 665675

Profit (RM/min)

Table 12 RGP prot values from experiments (HYSYS) and Taguchi method (ANOM) at optimal conditions. Case
1 2 3 4 5 6 7 8 9

2800 2500 2200 1900 1600 1300 100

150 200 Case #
8 7 9 5 4 6 2 1 3

HYSYS (RM/min)
1813.67 1840.20 1791.35 2009.59 2037.80 1981.95 2229.17 2259.73 2192.06

ANOM (RM/min)
1795.92 1831.21 1773.98 2021.39 2045.82 1982.43 2234.24 2265.63 2196.05

Deviation (%)
0.98 0.49 0.97 0.59 0.39 0.02 0.23 0.26 0.18

250

Time (min)
Fig. 5 RGP prots under optimal congurations for Cases 19.

are obtained for Cases 19, respectively. These values are compared with the ones obtained from Eq. (8), which are termed ANOM prot values. Eq. (8) contains two quantities, namely, global means and maximum differences of averages of factor k for the corresponding Cases m = 1,. . .,9. Assuming that HYSYS experimental results are the correct ones, the deviation from these values are termed error, Em for Cases m = 1,. . .9. The positive Em values indicate that HYSYS experimental results are lower than those obtained from ANOM. In this work, small Em values of less than 1% for all cases are obtained (Table 12). This means that optimal congurations of controllable factors for all the cases have been found. Optimal congurations of noise factors can be determined from the global means of prot in Cases 19. A minimum value of 1739.63 RM/min is obtained from Case 1 with the congurations H1 I1 . On the other hand, congurations H3 I2 in Case 8 yield a maximum value of 2216.94 RM/min. This shows that the optimum congurations of noise factors have been found in Case 8. If combined with the previous result, the optimal congurations of controllable and noise factors are A1 B1 C2 D3 E3 F3 G3 H3 I2 . When run in HYSYS, these congurations yield a prot value of 2259.73 RM/min. A comparison with ANOM result yields a small Em value of 0.3% conrming that this conguration is the optimum. In addition, the prot value obtained from running optimal conguration of factors is close to the maximum prot value of 2260.46 RM/min obtained from Run 3 in Case 8 with conguration A1 B1 C1 D3 E3 F3 G3 H3 I2 . Scrutinized closely, levels of factors are strikingly identical except for factor C (SRC102 output) that switches between levels 1 and 2. Resemblance of these two congurations draws an interesting point. Split-range controller SRC102 output may be varied between 90 and 100% without affecting much of the RGP prot. For SRC102 output values below 90%, a decrease by 1% in the controller output increases the demethanizer top temperature by about 0.2 C. This eventually leads to losses in ethane and propane products by about 0.90 and 2.25%, respectively. On the other hand, a controller output value above 100% is not allowed due to violation of physical constraint. In the analysis that uses SNR values, the conguration A1 B3 C2 D1 E3 F3 G2 seems optimal because the SNR values of each factor are the highest. Another run in HYSYS with this conguration coupled with H3 I2 conguration on noise factors yields a prot value of 2233.75 RM/min. This value is 1.2% lower than the one found based on the averaged prot analysis. When optimal congurations based on the average prot and the SNR values are compared, agreements are only found on levels of factors A, E and F. Differences in optimal levels of factors B, C, D and G are found to be responsible for lowering the prot. This result shows that an optimal conguration of factors cannot be determined from SNR response

plot. Instead, the optimal conguration can be found from the response plot based on the average prot values of all 9 cases as discussed earlier. When this conguration is run in HYSYS, the dynamic responses and the nal steady-state levels for Cases 19 can be plotted to illuminate several interesting points (Fig. 5). Highly uctuating trends in the early part of simulation reect difculties faced by the RGP to maintain stability due to changes in levels of all factors. After the initial rise in the prot values, a valley is clearly noticed after about 10 min of simulation effort for all 9 cases. For the last three cases, additional uctuations are noticed after 30 min of simulation time. In general, the valley phenomenon is caused by inadequacy of cooling for separation of natural gas liquids from the sales gas when the feed gas ow is raised from 250 to 280 ton/h. This corresponds to changes in factor H from level 1 to 2. Recall that the feed gas is cooled by increasing levels of factors A (SRC103 output), B (TC101 output) and C (SRC102 output). However, these factors are set at their lowest levels under the optimal conditions causing the prot values to initially drop. After occurrence of the valley, a sharp rise in the RGP prot margins is caused by increasing level of factor D (RC101 controller ratio) from level 1 to 3. Raising RC101 controller ratio from 0.003 to 0.15 increases the feed gas ow to the gas subcooled process section from about 1.1 to 34.2 ton/h. This action improves the recovery of valuable natural gas liquids and thus the RGP prot. In Cases 13 (Group I), prot trends produce the shallowest valley and settle at the lowest margins due to the low feed gas ow at 250 ton/h. Mild oscillations in Group I results are caused by changes in factors D, E, F, G from level 1 to 3 and variations in factor I (feed gas prices) at all three levels. On the other hand, Cases 46 (Group II) generate the deepest valley but settle above Group I steady-state values. Higher settling trends of Group II prot are caused by the increase in the feed gas ow from 250 to 280 ton/h. Raising the feed gas ow also causes the valley occurrence due to insufcient cooling of additional 30 ton/h of the feed gas. In Cases 79 (Group III), the major uctuation of prot trends is seen after 30 min of simulation time. This phenomenon is attributed to the equivalent delay in switching time of level of factor D (RC101 controller ratio) from levels 1 to 3. The delay is required to prevent instability of the RGP dynamic model when changes on other factors are imposed simultaneously. Trends of Group III settle at the highest levels as compared with trends of the other two groups. Results from this observation indicate that factor H (plant load) at level 3 yields the highest prot margins. Slight differences of settling trends are caused by variations in levels of factor I. The highest prot margin obtained in Case 8 reveals that I2 (FG-B with

chemical engineering research and design 8 9 ( 2 0 1 1 ) 665675

675

moderate carbon dioxide content) conguration is the optimum although price of FG-B is in between those of FG-A and FG-C.

for validation. In all the Cases 19, results from HYSYS experiments are compared against those from ANOM. Remarkable agreements found in all cases verify the optimality of factorial conguration.

5.

Conclusions Acknowledgement
This work is fully sponsored by the Universiti Teknologi Petronas (UTP).

Signicance of 7 controllable and 2 noise factors affecting the RGP prot is studied by conducting 243 experiments in a crossorthogonal-array set up. Three controllable factors handling the RGP energy consumption top the ranked list. A combined effect of factors A (SRC103 output) and B (TC101 output) controls the entire plant temperature. Higher values of these factors cool the plant and thus enhancing recovery of natural gas liquids. In a balancing move, energy intake through factor G (TC102 output) is vital in ensuring smooth separation of the feed gas in the demethanizer. The RGP prot can also be improved by setting higher level of factor E (FC104 output). This action increases the amount of hydrocarbon injection to the sales gas product, and thus boosting the sales gas gross heating value and ow. Contributions from factors C (SRC102 output), D (RC101 controller ratio) and F (PC101 set point) are found to be less signicant in optimizing the RGP prot. These controllable factors play more important roles in minimizing the effects of noise factors as shown in ANOM and ANOVA results of SNR values. Among the two noise factors, factor H (plant load) is more signicant than factor I (feed gas prices). Higher plant load increases the amount of feed gas to be processed, which in turn produces more sales gas and natural gas liquids products. Maximum RGP prot is derived from an optimal conguration of both controllable and noise factors. This unique conguration of low, medium or high levels of individual factors was selected based on a response plot of averaged prot values. The optimal conguration of factors is set in HYSYS

References
Cheng, W.T., Li, H.C., Huang, C.N., 2008. Simulation and optimization of silicon thermal CVD through CFD integrating Taguchi method. Chemical Engineering Journal 137, 603613. Chiang, K.-T., 2005. Optimization of the design parameters of parallel-plain n heat sink module cooling phenomenon based on the Taguchi method. International Communications in Heat and Mass Transfer 32, 11931201. Engin, A.B., Ozdemir, O., Turan, M., Turan, A.Z., 2008. Color removal from textile dyebath efuents in a zeolite xed bed reactor: determination of optimum process conditions using Taguchi method. Journal of Hazardous Materials 159, 348353. Lee, K., Kim, J., 2000. Controller gain tuning of a simultaneous multi-axis PID control system using the Taguchi method. Control Engineering Practice 8, 949958. Roy, R.K., 1990. A Primer of Taguchi Method. van Nostrand Reinhold, New York. Taguchi, G., Konishi, S., 1987. Orthogonal Arrays and Linear Graphs. American Supplier Institute, Inc., Dearborn, MI. Yang, K., Teo, E.-C., Fuss, F.K., 2007. Application of Taguchi method in optimization of cervical ring cage. Journal of Biomechanics 40, 32513256. Yusoff, N., Ramasamy, M., Yusup, S., 2008. A simulation study on dynamics and control of a refrigerated gas plant. In: Proceedings of the FOCAPO-5, Boston, USA, pp. 263266.