Академический Документы
Профессиональный Документы
Культура Документы
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/251604637
CITATIONS READS
10 249
4 authors, including:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Gholamreza Zahedi on 01 April 2015.
a r t i c l e i n f o a b s t r a c t
Article history: The main objective of this study is modeling and optimization of an industrial Hydrocracker Unit (HU) using
Received 8 April 2011 Articial Neural Network (ANN) model. In this case some data from an industrial hydrocracker plant were
Accepted 26 July 2011 collected. Two-thirds of the data points were used to train ANN model. Among the various networks and
Available online 9 August 2011
architectures, two multilayer feed forward networks with Back Propagation (BP) training algorithm were
found as the best model for the plant. Inputs of both ANNs include fresh feed and recycle hydrogen ow rate,
Keywords:
Hydrocracker unit
temperature of reactors, mole percentage of H2 and H2S, feed ow rate and temperature of debutanizer,
Renery pressure of debutanizer receiver, top and bottom temperature of fractionator column and pressure of
Modeling fractionator column. The rst network was employed to calculate the specic gravity of gas oil, kerosene, Light
Articial neural network Naphtha (LN), Heavy Naphtha (HN), gas oil and kerosene ash point and gas oil pour point. The second
network was used to calculate the volume percent of C4, LN, HN and kerosene, gas oil and fractionators
column residual (off test). Unseen data points were used to check generalization capability of the best
network. There were good overlap between network estimations and unseen data.
In the next step of study sensitivity analysis was carried out on plant to check the effect of input variables on
the plant performance. In this case temperature was found as the most affecting parameter in the plant.
Finally optimization was performed to maximize the volume percent of gas oil, kerosene, HN and LN
production and to identify the sets of optimum operating parameters to maximize these product yields.
Optimum conditions were found as feed ow rate of 113.2 m3/h, reactor temperature of 413 C, hydrogen
ow rate of 111.3 MSCM/h and LN () feed vol.% of 9.22.
2011 Elsevier B.V. All rights reserved.
0920-4105/$ see front matter 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.petrol.2011.07.019
628 I. Alhajree et al. / Journal of Petroleum Science and Engineering 78 (2011) 627636
Fig. 2. Schematic representation of the multilayer feed forward neural network for (a) the rst network (SCG) and (b) the second network (CGF).
630 I. Alhajree et al. / Journal of Petroleum Science and Engineering 78 (2011) 627636
Fig. 3. Comparison of measurement and simulated results of the rst ANN model For (a) pour point of gas oil, (b) ash point of gas oil, (c) ash point of kerosene, (d) specic gravity
of gas oil, (e) specic gravity of kerosene (f) specic gravity of HN and (g) specic gravity of LN.
is determined by the connections between elements. An ANN can be train a network. NNs can be trained to solve problems that are difcult
trained to perform as a particular function by adjusting the values of for conventional computers or human beings (Choudhury et al., 1999;
the connections (weights) between elements, so that a particular Pastor-Brcenas et al., 2005; Kucuk and Derebasi, 2006).
input leads to a specic target output. The network is adjusted, based ANNs are strong tools for complex and exible input-output
on a comparison between the output and the target, until the network mappings (Pastor-Brcenas et al., 2005). Construction of a neural
output matches the target. Many input-target pairs are needed to network includes the determination of the network properties (the
I. Alhajree et al. / Journal of Petroleum Science and Engineering 78 (2011) 627636 631
Fig. 4. Comparison of measurement and simulated results of the second ANN model For the volume percent of (a) C4, (b) LN, (c) HN, (d) kerosene, (e) gas oil and (f) off test.
Fig. 5. The effect of temperature on the volume percent of (a) gas oil, (b) kerosene, (c) HN and (d) LN.
1999).The error between the actual and predicted outputs is used to 4. ANN modeling
modify the weights. This training method which uses the set of input-
output (examples) is known as supervised training. Multi-layer 4.1. Industrial data sets and modeling procedure
perception (MLP) is the most popular feed forward network for
prediction, so that all signals ow in a single direction from the input To build up an ANN for simulating the HU, necessary data were
to the output (Kucuk and Derebasi, 2006; Reddy et al., 2006). provided. Among the available data from Tabriz renery HU, 108 data
In MLPs, the back propagation (BP) algorithm for supervised were screened based on principle component analysis (Reddy et al.,
training is often adopted to propagate the error from the output 2006). MATLAB software was employed to build the ANN models
neurons and modify the weight for the neurons in the hidden layers (Demuth and Beale, 2010).
(Kucuk and Derebasi, 2006). Various architectures of MLP, RBF and BP were employed to model
The error gradient (wki) for the kth output or hidden node is the unit. For scaling purposes, each type of input and output data were
given by (Kucuk and Derebasi, 2006). normalized between 0 to 1. Each ANN was trained with 2/3 of data set
and the remaining 1/3 of samples was employed for testing the
p obtained ANN.
wki = ak xki 1
In order to nd the optimum number of hidden layers and nodes
p within each layer trial and error approach was used. The trial and
where, a, xki are learning rate and kth node input pattern respectively.
error approach includes one, two and three hidden layers, and 40 to
k is hidden layer kth node error:
120 nodes per each hidden layer. Tables 1 and 2 show HU variables
and their operating range.
k = ak j wj k 2
j = Ik After the trial and error it was found that a single ANN can
simulate the plant very well. Therefore based on our previous
experience (Zahedi et al., 2008), two ANNs in parallel were used to
For output node error can be calculated based on the following
model the plant.
equation:
In the rst ANN, a BP network with the SCG algorithm was found as
p p
the best network for calculation of the specic gravity of gas oil,
k = ak tk yk 3 kerosene, LN, HN and the ash point of gas oil and kerosene and gas oil
pour point (seven output vector).
p p
where tk is kth node target pattern, yk is kth node output pattern In the next ANN, the volume percent of C4, LN, HN, kerosene, gas oil
(Kucuk and Derebasi, 2006). and residual fractionator column (off test) (6 output vectors) were
I. Alhajree et al. / Journal of Petroleum Science and Engineering 78 (2011) 627636 633
Fig. 6. The effect of total feed ow rate (make up and recycle) on the volume percent of (a) gas oil, (b) kerosene, (c) HN and (d) LN.
calculated using BP network with Fletcher-Powell Conjugate Gradient 4.2. ANN modeling results
(CGF) algorithm. Both networks had the same inputs (20 input).
It is important to nd the optimum number of nodes in the hidden As mentioned before, two ANN models were developed for HU
layer which provide good estimates of the outputs. The rst network modeling. The generalization results of the rst and second networks
was designed with 74 and 74 neurons in the rst and the second are illustrated in Figs. 3 and 4, respectively. These gures show the
hidden layers, respectively, while, in the second ANN, there were ANN outputs versus the number of unseen industrial data.
three hidden layers with 61, 71 and 61 neurons, respectively. These gures show good capability of both ANNs for estimation of
In training the networks, the sigmoid and simple linear activation unseen data. Mean relative percent of error between unseen plant
functions were used in the hidden layers and output layers: data and network outputs was obtained as 1.3924%. Details are
summarized in Table 3. According to Table 3, the proposed ANN
1 models can model the unseen data with excellent precision and can be
f x = Logistic sigmoid function 4 employed for optimization and further assessment of the plant.
1 + ex
Fig. 7. The effect of total ow rate of inlet H2 (make up and recycle) on the volume percent of (a) gas oil, (b) kerosene, (c) HN and (d) LN.
simulator. Sensitivity analysis of a system was conducted using the trained from the trained ANN model by observing the change in the value of the
NNs to characterize the relationship between inputs and outputs. The kth output brought by a small change in the value of the jth input at any
sensitivity of the outputs to the inputs dened by Eq. (7) was estimated operating point (Reddy et al., 2006).
At each operating point P, the j th input was varied over its entire
Optimum operating parameters to range, and the corresponding variation in the kth output can be
maximize LN (volume %) plotted to produce a sensitivity curve (Reddy et al., 2006). In this way,
the contribution of that variable to the network response is excluded
(Aminian and Shahhosseini, 2008).
9.4 Using the second designed network, the effect of the operating
parameters of HU, like temperature of reactors, volume ow rate of
9.2
feed (make up and recycle) and volume ow rate of hydrogen (make
9 up and recycle), on volume percent of main outputs including gas oil,
kerosene, LN and HN at reactors pressure of 185 kg/cm 2 were studied.
8.8
Table 4 shows domain of change for operating parameters.
Light Naphtha %
8.6 Fig. 5 illustrates the effect of temperature on the gas oil, kerosene,
HN and LN percent while feed ow rate and inlet hydrogen ow rate
8.4
are set in their average values. As shown in this gure, temperature is
8.2 the most affecting input variable. As reactor temperature changes from
380 to 440C, the trend of the volume percent of gas oil, kerosene, HN
8
and LN changes according to Fig. 5 (a)(d), respectively.
7.8
Temperature+300
7.6 Hydrogen*1000 Table 5
Feed Optimum parameters.
7.4
100 105 110 115 120 125 130 Feed Temperature of reactors Hydrogen rate LN
Temperature, Hydrogen and Feed (m3/h) (C) (m3/h) (vol.%)
Edgar, T.F., Himmelblau, D.M., Lasdon, L.S., 2001. Optimization of Chemical Processes, pruning techniques in neural networks for surface ozone modeling. Ecol. Model.
second ed. McGraw Hill, New York. 182, 149158.
Elkamel, A., Al-Ajmi, A., Fahim, M., 1999. Modeling the hydrocracking process using Reddy, N.S., Lee, C.S., Kim, J.H., Semiatin, S.L., 2006. Determination of the beta-approach
articial neural networks. Petrol. Sci. Technol. 17 (910), 931954. curve and beta-transus temperature for titanium alloys using sensitivity analysis of
Falla, F.S., Larini, C., Le Roux, G.A.C., Quina, F.H., Moro, L.F.L., Nascimento, C.A.O., 2006. a trained neural network. Mater. Sci. Eng. A. 434, 218226.
Characterization of crude petroleum by NIR. J. Petrol. Sci. Eng. 51, 127137. Sullivan, R.F., Boduszynski, M.M., Fetzer, J.C., 1989. Molecular transformations in
Jaiswal, S., Benson, E.R., Bernard, J.C., Van Wicklen, J.L., 2005. Neural Network Modelling and hydrotreating and hydrocracking. Energy Fuels. 3 (5), 603612.
Sensitivity Analysis of a Mechanical Poultry Catching System. Biosystems Eng. 92 (1), Tabriz, Renery, 2009. hydrocracker unit data.
5968. Shirvani, Yazdan, Zahedi, Gholamreza, Bashiri, Mohsen, 2010. Estimation of Sour
Kucuk, I., Derebasi, N., 2006. Sensitivity analysis for estimation of power losses in Natural Gas Water Content. J. Pet. Sci. Eng. 73, 156160.
magnetic cores using neural network. J. Phys. Chem. Solids. 67, 24732477. Zahedi, G., Lohi, A., Karami, Z., 2009. A neural network approach for identication and
Lyons William, C., 2005. second edition. Standard Handbook of Petroleum and Natural modeling of delayed coking plant. Int. J. Chem. Reactor Eng. 7, A16.
Gas Engineering, Volume 2. GULF PROFESSIONAL PUBLISHING, USA. Zahedi, G., Mohammadzadeh, S., Moradi, G., 2008. Enhancing gasoline production in an
Maples, R.E., 2000. Petroleum Renery Process Economics, second edition. industrial catalytic-reforming unit using articial neural networks. Energy Fuels.
Motlaghi Jalali, F., Nili Ahmadabadi, M., 2008. An expert system design for a crude oil 22, 26712677.
distillation column with the neural networks model and the process optimization Zahedi, G., Fgaier, H., Jahanmiri, A., Al-Enezi, G., 2006. Articial neural network identication
using genetic algorithm framework. Expert Systems with Applications. 35, and evaluation of hydrotreater plant. Petrol. Sci. Technol. 24, 14471456.
15401545. Zahedi, G., Parvizian, F., Rahimi, M.R., 2010. An expert model for estimation of
Pastor-Brcenas, O., Soria-Olivas, E., Martn-Guerrero, J.D., Camps-Valls, G., Carrasco- distillation sieve tray efciency based on articial neural network approach.
Rodrgueza, J.L., del Valle-Tascn, S., 2005. Unbiased sensitivity analysis and Journal of Applied Science 10( (12), 10761082.