Keywords 
Artificial neural networks model; Neural networks
predictive control; NARMA controller; Crude oil distillation unit 
Nomenclature 
E Error criteria for network convergence 
G(s) Transfer function 
N_{u} Horizons over tracking error 
t Time, (min) 
T Temperature, (°C) 
u(k) System input 
Wij Weight value between input and hidden layer 
X Input of neuron 
y Output of neuron 
y_{r} Desired response 
Δ Difference 
α Momentum rate 
η Learning rate 
b Bias 
ρ Weighting parameter 
Introduction 
The crude oil distillation unit (CDU) fractionation column
separates the feed crude into different cuts suitable for the different
refinery processing units. CDU today is facing new challenges in order
to meet the requirements with respect to improve fuel properties,
product quality and increase the yields of the distillate products
with meeting environmental laws. A lot of crude units currently
operate with different feed slates to their original feed specifications
to satisfying the demands of the market. Most petroleum distillates,
especially those from atmospheric distillation tower, have different
physical properties depending on the characteristics of the crude oil
[1]. The scope of the control systems in process industries has been
broadened from the basic regulatory control to advanced control strategies. The temperature control is based on the assumption that the
product composition can satisfy its specification when an appropriate
tray temperature is kept constant at its setpoint [2]. In the control
of crude oil distillation columns is usually difficulty to get accurate
and reliable product composition measurements without time delay.
Many composition analyzers such as gas chromatography, NIR (Near
Infrared) analyzers, suffer from large measurement delays and high
investment and maintenance costs and usually possess significant time
lags. The overall time lags in composition measurements are typically
between 10 to 20 minutes. Also in inferential control of product
composition is used by estimation from other measured variables [3]. 
Nunzio et al. [4] presented the neural networks controller (NNC)
for quality predictions of a crude unit. The Neural networks (NN)
implemented compare between predicted and measured quality
on the light gasoil stream over a five week. The average % absolute
error is about 1.23 with a standard deviation of 1.1. Ali and Khalid
[5] implemented intelligent control technology of NN for crude
fractionation tower. The simulation results for modeling of the column
were used to control the several products property such as naphtha
95% cut point and naphtha Reid vapor pressure. The sum squared error
goal of training for control is 0.1 while for in the verification mode is
0.097. Lee et al. [6] used artificial neural network (ANN) controller to
identify the feed and product. Two ideas are used to identify the control
on feed characteristic as a real time basis. The proposed method can
be effectively used for controlling process optimization. Pavel et al. [7]
upgraded ANN controller for online inferential property estimation.
The neural networks run as function blocks within the automation
system’s controller. Pasadakis et al. [8] introduced ANN controller. The properties of interest distillation curve with 5% recovery and cold
properties of diesel fuel were measured experimentally. Result shows
that it can be easily employed in the refinery environment for online
process control. Omole et al. [9] developed back propagation neural
network (BPNN) model for predicting and control crude oil. The data
points of ANN models performed better than the existing empirical
correlations to control crude oil viscosity. 
Henri and Olatunbosun [10] developed NN controller for a crude
oil distillation column, field data used from a working unit of crude
oil. They estimated the correlation coefficients between the obtained
values from NNC and the field values for steam flow and three streams
of reflux flow. Lekan et al. [11] presented a simulation of crude oil
distillation column and applied the ANN controller. Good results
accuracy is obtained and the deviation between NNC from each other
of 1.98%. Richalet [2] worked on model based predictive control for
crude oil unit. The results showed that the good performance and
robustness could be obtained under wide operating condition. Sharad
and James [12] proposed inferential measurement to control variable
by the systematic approach using neural networks controller. They
developed a correlation to predict the ASTM 95% endpoint of kerosene
with an error standard deviation of 1.7°C. Aboujeyab et al. [13] used
model predictive control (MPC) of a crude oil distillation column. The
results showed that 2.5% increase in production rate and 0.5% increase
in product recovery. Kemaloglu et al. [14] reviewed the application of
a model predictive controller algorithm to a crude oil unit. The system
responses for different changing in setpoint in the product qualities
to be increase 11% in the kerosene yield was achieved as a result of
decrease in naphtha yield. 
Volk et al. [15] tested the multivariable predictive control of an
crude distillation column. This controller keeps the setpoints of
the distillate and bottom concentration constant. The linear control
algorithm is valid in the vicinity of the working point for changes of
about 10%. Gabriele et al. [16] introduced simulation model of CDU
and controlled by a multivariable predictive controller. The controller
tuning is implemented on the actual plant and carry out closed loop
identification tests from which the predictive controller implemented.
Aliyev et al. [17] tested the crude refinery unit control by using two
type of control architecture. The nonlinear model predictive (NMPC)
was able to track set points and response of NMPC is better than of
response PI control. 
Haydary and Tomas [18] investigated two different control
methods based on composition of ASTM D86 95% boiling point and
temperature for real crude oil distillation plant. Experimental ASTM
D86 curves of different products were compared to those obtained by
simulations. Sampath [1] used a control layers in the AspenHYSYS
simulator for crude oil distillation unit, the first control layer is PID
and the second control layer is MPC. He concluded that the MPC
can handle constraints and presents good robustness features against
model mismatch and perturbations. Goncalves et al. [19] applied PID
controller for atmospheric distillation unit of crude oil refinery. The
dynamic model is developed and combined with a suitable control for
several process operating conditions. They study the step responses
for quality specifications like ASTM D86 95%, and production flow
changes. Rogina et al. [20] worked on light naphtha in CDU control.
Experimental and laboratory analyses data was used for neural network
based model. The analyses show that conditions at the column top
temperature most affect the RVP and NNC are acceptable result for
RVP estimation. Mohammadi et al. [21] investigated simulation and
control for Kermanshah refinery by applying the PID controller. The behavior controller was observed after the changing of the crude oil
feed flow rate by 3% for five hours. The results show that temperature
controllers are faster and more sensitive than the other controllers. 
The purpose of this paper builds the artificial neural network model
for crude oil distillation unit, and applies neural networks predictive
and NARMAL2 controller to the crude oil distillation column. 
Neural Networks Predictive Controller 
The neural network predictive controller calculates the control
input that will optimize plant performance over a specified future time
horizon. The first step of model predictive control is training a neural
network to represent the forward dynamics of the plant. The prediction
error between the plant output and the neural network output is used
as the neural network training signal. The neural network plant model
uses previous inputs and previous plant outputs to predict future values
of the plant output. The model predictive control method is based on
the receding horizon technique. The neural network model predicts the
plant response over a specified time horizon. The predictions are used
by a numerical optimization program to determine the control signal
that minimizes the following performance criterion over the specified
horizon. 
(1) 
Where N_{1}, N_{2}, and N_{u} are the horizons over which the tracking
error and the control increments are evaluated. The u′ variable is the
tentative control signal, y_{r} is the desired response, and y_{m} is the network
model response. The ρ value determines the contribution that the sum
of the squares of the control increments has on the performance index.
The optimization block determines the values of u′ that minimize J,
and then the optimal u is input to the plant. Figure 1 shows the Block
diagram of the neural network predictive control. 
Nonlinear Autoregressive Moving Average NARMAL2
(Feedback Linearization) Controller 
The central idea of this type of control is transforming nonlinear
system dynamics into linear dynamics by canceling the nonlinearities.
As with model predictive control, the first step in using feedback
linearization (or NARMAL2) control is identifying the system to be
controlled and then choose a model structure to use. The nonlinear
autoregressivemoving average (NARMA) model is used to represent
general discretetime nonlinear systems as expressed in Equation. 2. 
(2) Output Response 
Where u(k) is the system input, and y(k) is the system output. For the
identification phase, a neural network is trained to approximate the
nonlinear function N. This is the identification procedure used for
the NN predictive controller. The system output is equal to reference
trajectory (y (k + d) = y_{r} (k + d)). 
The next step is developing a nonlinear controller as the following
form: 
(3) 
The resulting controller would have the following form 
(4) 
Using Equation. 4 directly can cause realization problems, because the
control input u(k) must is determined based on the output at the same
time, y(k). Using the NARMAL2 model, you can obtain the controller 
(5) 
which is realizable for d ≥ 2. Figure 2 shows the Block diagram of the
NARMAL2 control [22]. 
Simulation Work 
The neural network architecture for the design of the crude oil
distillation column (CODC) consists of thirteen inputs with one
hidden layer which consist of nine nodes and six outputs making a
total of 34 nodes distributed over the three layers. The inputs to the
network are volumetric flow rates of top pumparounds, intermediate
pumparounds, bottom pumparounds, steam, reflux, naphtha, kerosene,
and light gas oil (LGO) toptemperature, toppressure, specific gravity,
temperature, and volumetric flowrate of feed. The outputs from the NN
architecture are temperatures of naphtha, kerosene, LGO D86 95%,
top, intermediate and bottom pumparounds. The back propagation
algorithm is used for the ANN of crude oil distillation unit. Figure
3 shows the neural network architecture of the crude oil distillation
column, 1487 records set are collected from the designed unit in aspen
HYSYS simulator, these data are collected by making step changed in
the manipulated variables for dynamic case and record the response for
each input and output mentioned earlier in aspenHYSYS simulator
and converted to Excel spreadsheet, the range of data used in the
training is shown in Table 1. These data are used in MATLAB simulator
to build NN model for CDU. Nonlinear autoregressive network with
exogenous inputs (NARX) are used for the ANN model in MATLAB. Since the case is studying the dynamic behavior and control of crude
oil distillation unit. Input and output data are loaded to the workspace
from excel spreadsheet. 70% of the data are selected for training and
30% are used for validation and testing. The simulator normalized the
training data between (1,1). 
After establishing the ANN model that has been developed and
converted to simulink, the control system is built for this model using
neural network predictive and NARMAL2 control methods. The
neural network controller that is implemented uses a neural network
model of a nonlinear plant to predict future plant performance. NN
predictive controller and NARMAL2 controller blocks are installed
to the SIMULINK window and connected with NN model and filling
the controller’s parameters then identified the plant. Controllers are
trained by train algorithm. We applied six NN predictive controller
and six NN NARMAL2 controllers to control each one of the six
model outlet. The controller architecture for the six controlled variable
are shown in the Figures 49. 
Results and Discussion 
Model of crude oil distillation column 
Artificial neural network model is established successfully for the
crude oil distillation column. Figure 10 shows neural network training
performance. This figure shows the architecture of the network build
and the value of mean square error (MSE) is 0.25 and the iteration time
for the program is 18 iterations. Figure 11 shows the neural network
validation performance is equal to 23.34 at epoch 12 (time steps
for adaption). Neural network training regression for training and
validation are 0.99996 and 0.99844 respectively as shown Table 2. This
model has been tested with different step changes in input variables
and we get satisfied result for the output but the step changes should
be in the training limits. Neural network model for the nonlinear unit
of crude distillation unit promise a good performance to handle the
complexity and nonlinearity at the same time this is due to the full
representation of the nonlinear dynamic of the unit. 
Neural networks controller 
The control system is desigend for this model using neural
network controllerby using two methods, neural network predictive
and NARMAL2. The controlled variables are temperatures of top
pumparound, intermediate pumparound, bottom pumparound,
naphtha D86, kerosene D86 and LGO D86. A step change is made in
feed temperature to test the controller as shown in Table 2. Also in these
runs the changes in temperatures are within the limit of the training of
the neural network due to the inherent property of neural networks and
it is good in interpolation but is not good in extrapolation. Advantages
of NN based controllers do not require any tuning of the control
parameters also can take care of a nonlinear model of the process,
compute the manipulated variables rapidly and produce less oscillation.
Finally has less offset value and neural controller has more suitable.
The temperature response reach the steady state value in less time
and neural controller has lower overshoot. Artificial neural network
controller learns system and it has got generalization capabilities. The
controllers based on this neural network model are able to take into
account any significant process model mismatch. A step change in feed
temperature from 340 to 350°C is carried out using neural network
predictive and neural network NARMAL2 controller to control the
temperatures of naphtha, kerosene, LGO D86 95%, top, intermediate
and bottom pumparounds. Figures 1217 show the controller’s
behavior for naphtha, kerosene, LGO D86 95%, top, intermediate, and
bottom pumparounds return temperatures. The comparison between
the behavior of NN predictive and neural NARMAL2 controllers are
made by using mean square error criterion. It can be seen from the
Table 2 the MSE for NARMAL2 controller is less from the obtained
of neural predictive controller. The maximum MSE for NARMAL2 is
103.1 while it is 182.7 for neural predictive. Both of these controllers
are able to eliminate the offset without any overshoot. The satisfactory
performance is due to the full representation of the nonlinear
dynamics of the crude oil distillation column. The NARMAL2
controller responds as quickly as NN predictive. They indicate that the
NARMAL2 give smallest overshoots, shortest settling times and shows
less oscillation, smoother and better control performance than the NN
predictive controllers with smaller MSE error values acquired. 
Conclusion 
The results presented in this paper have clearly shown the ability
of neural networks to act as process controllers. The ANN is advance
method can be used to model any nonlinear, complex and multivariable
system. It gives the desired performance if trained well, for this study
the ANN are useful tool for representing and predicting the plant
output for specific input. Also the results have shown priority of neural
network NARMAL2 controller in crude oil distillation column. From
simulation results which give a less offset value and the temperature
response reach the steady state value in less time with lower overshoot
compared with neural network predictive controller. Finally the MSE
of NARMAL2 is less than MSE of neural network predictive control. 
References 
 Sampath Y (2009) Framework for operability Assessment of production facilities: An Application to A primary Unit of A crude oil refinery. MSc Thesis, Department of Chemical Engineering, Louisiana State University, USA.
 Richalet J (1993) Industrial Application of Model Based Predictive Control. Pergamon Press Ltd., Automatica 29: 12511274.
 Jie Z (2001) Inferential Feedback Control of distillation composition based on PCR and PLS Model. Proceedings of the American control conference. pp: 11961201.
 Nunzio B, Riccardo M, Giuseppe R, Michele D (2000) Neural NetBased Inferential Quality control of A crude unit. Hydrocarbon Engineering. pp: 3338.
 Ali Z, Khalid B (2001) Application of neural networks in oil refineries. Ras Tanura refinery, Saudi Arabia, CRC Press LLC.
 Lee Doug H, Euy SL, Sang JP (2006) Characterization of crude feed and products from operating conditions by using continuous probability functions and inferential models. Korean J Chem Eng 23: 517521.
 Pavel K, Wulf S, Alexei S (2006) Using neural network technology for virtual sensing in crude refining units: ISA Expo. Process Analysis & Modeling Consultant to Lukoil Moscow.
 Pasadakis N, Sourligas S, Foteinopoulos C (2006) Prediction of the distillation profile and cold properties of diesel fuels using midIR spectroscopy and neural networks. Fuel J 85: 11311137.
 Omole O, Falode O, Deng A (2009) Prediction of Nigerian crude oil viscosity using artificial neural network. Petroleum & Coal J 3: 181188.
 Henri EZ, Olatunbosun A (2010) Neural network controller for A Crude oil distillation column: Asian Research Publishing Network (ARPN). Journal of Engineering and Applied Sciences 5: 7482.
 Lekan TP, Gutti B, Alfred AS (2013) Expert system design and control of crude oil distillation column of a Nigerian refinery using artificial neural network model. IJRRAS 15: 337346.
 Sharad B, James W (2001) Development of Inferential measurements using neural networks. ISA Transactions 40: 307323.
 AbouJeyab RA, Gupta YP, Gervais JR, Branchi PA (2001) Constrained multivariable control of a distillation column using a simplified model predictive control algorithm. Journal of process control 11: 509517.
 Kemaloglu S, Ozgen EK, Dila G, Ozgur C (2004) Model predictive control of a crude distillation unit an industrial Application. Turkish Petroleum Refineries Corporation, Izmit. p: 41780.
 Volk U, Kniese DW, Hahn R, Haber R, Schmitz U (2005) Optimized multivariable predictive control of an industrial distillation column considering hard and soft constraints. Control Engineering Practice 13: 913927.
 Gabriele P, Lorenzo G, Alessandro B, Gabriele M, Filippo T (2006) Rigorous Simulation and Model predictive control of A crude Distillation Unit. International symposium on advanced control of chemical processes. ADCHEM. pp: 635640.
 Aliyev TL, Stamps AT, Gatzke EP (2008) Improved Crude Oil Processing Using SecondOrder Volterra Models and Nonlinear Model Predictive Control. American Control Conference. pp: 22152220.
 Haydary J, Tomas P (2009) Steadystate and Dynamic Simulation of crude oil distillation using Aspen Plus and Aspen dynamics. Petroleum and Coal 51: 100109.
 Goncalves DD, Fernando GM, Azevedo SF (2010) Dynamic Simulation and Control: Application to Atmospheric Distillation Unit of Crude Oil Refinery. 20th European Symposium on Computer Aided Process Engineering, Portugal.
 Rogina A, Isko I, Mohler I, Ujevic Z, Bolf N (2011) Soft sensor for continuous product quality estimation (in crude distillation unit). Chemical engineering research and design. pp: 20702077.
 Mohammadi A, Shahraki S, Sadeghi J (2012) Simulation, control and sensitivity analysis of crude oil distillation unit. Journal of Petroleum and Gas Engineering 3: 99113.
 Mark HB, Martin TH, Howard BD (2013) Neural Network Toolbox™ User’s Guide.
