JID: JTICE

ARTICLE IN PRESS [m5G;January 23, 2018;13:11]

Journal of the Taiwan Institute of Chemical Engineers 0 0 0 (2018) 1–11

Contents lists available at ScienceDirect

Journal of the Taiwan Institute of Chemical Engineers

journal homepage: [Link]/locate/jtice

A data-driven soft-sensor for monitoring ASTM-D86 of CDU side

products using local instrumental variable (LIV) technique
Bahareh Bidar, Mir Mohammad Khalilipour, Farhad Shahraki, Jafar Sadeghi∗
Center for Process Integration and Control (CPIC), Department of Chemical Engineering, University of Sistan and Baluchestan, Zahedan 98164, Iran

a r t i c l e i n f o a b s t r a c t

Article history: Atmospheric crude distillation unit is the main unit operation in petroleum refining industries. The main
Received 5 September 2017 difficulties in quality control of column are the availability of quality measurements. The design of prod-
Revised 2 December 2017
uct quality estimator will help improve quality monitoring and control performance in oil refinery indus-
Accepted 7 January 2018
try by accurately predicting the side products properties, simultaneously. The objective of this paper is
Available online xxx
to design and implement state dependent parameter (SDP) based soft sensors using local instrumental
Keywords: variables (LIV) technique for an industrial atmospheric crude distillation unit. On the basis of tray tem-
Data-driven soft sensor perature measurements of the column, soft sensor models for estimation of 95%ASTM-D86 of product
Instrumental variable streams have been developed. Three soft sensors are separately designed in an offline manner for each
State dependent parameter product quality with steady-state data of the column. The performance of proposed soft sensors is evalu-
Quality prediction ated through testing data and also by online implementation in simulated control system. The prediction
ASTM-D86 temperature
results, after tuning controller parameters, show excellent agreement with quality predictions from the
Crude distillation column
rigorous model. Based on developed soft sensors, it is possible to estimate product properties in a con-
tinuous manner with minimum delay compared to laboratory ASTM analysis and apply perfect control as
well as compliance with product quality specifications.
© 2018 Published by Elsevier B.V. on behalf of Taiwan Institute of Chemical Engineers.

1. Introduction ing the product properties within the same boiling interval by ex-
ternal factors (e.g. feed characteristics) can result in non-uniformity
Crude distillation unit (CDU) is one of the most important units of product quality. Thus, an inferential sensing-based control strat-
in the refineries, which separates the preheated crude oil into re- egy is tenable, which needs less manual effort and maintenance
spective product fractions like naphtha, kerosene and gas oil, etc. cost [2,3].
The stringent quality control requirement in a highly competitive The soft sensor is a key technology to infer the important qual-
market, makes it essential that all the necessary product proper- ity variables, which are difficult-to-measure online. However, the
ties such as Reid vapor pressure (RVP) for volatile products, flash- soft sensor is accurate enough; the predicted qualities can then
point for light distillates, pour point for heavier fractions, etc. are be used as a feedback for automatic control and optimization pur-
monitored online and kept under control. These product properties poses. However, there are still many problems with the existing
are generally not available online and usually measured in an of- estimators that require the development of new techniques. Nowa-
fline manner with intervals of 8–24 h, which may lead to improper days, data-driven soft sensors such as partial least squares (PLS)
control performance. Therefore, the product properties are con- [4,5], artificial neural networks (ANN) [6,7], support vector regres-
ventionally controlled using the range of boiling points. There are sion (SVR) [8,9] have gained much popularity in the industrial pro-
three types of boiling point analysis, namely ASTM1 -D86 (Engler), cesses.
ASTM-D158 (Saybolt) and true boiling point (TBP). The ASTM-D86, In relation to the use of data-driven soft sensors for esti-
among the methods, is the standard test method for distillation of mation of product properties in CDU, many studies have been
petroleum products at atmospheric pressure [1]. However, chang- done in recent decades. To handle the strongly correlated process
variables in CDU, the principal component analysis (PCA) and
PLS approaches have been used. Wang et al. [10] developed a
∗
Corresponding author.
PLS-based soft sensor and applied to an industrial CDU. The ASTM
E-mail addresses: [Link]@[Link] (B. Bidar), [Link]@[Link]
(M.M. Khalilipour), fshahraki@[Link] (F. Shahraki), sadeghi@[Link] (J.
90% distillation temperature (D90) of product streams and 14 pro-
Sadeghi). cess variables are considered as the quality index and predictors,
1
American Society for Testing Materials. respectively. Nevertheless, the PCA and PLS can only extract linear

[Link]
1876-1070/© 2018 Published by Elsevier B.V. on behalf of Taiwan Institute of Chemical Engineers.

Please cite this article as: B. Bidar et al., A data-driven soft-sensor for monitoring ASTM-D86 of CDU side products using local instru-
mental variable (LIV) technique, Journal of the Taiwan Institute of Chemical Engineers (2018), [Link]
JID: JTICE
ARTICLE IN PRESS [m5G;January 23, 2018;13:11]

2 B. Bidar et al. / Journal of the Taiwan Institute of Chemical Engineers 000 (2018) 1–11

eral drawbacks such as getting stuck in local minima, weak inter-

Nomenclature pretation capability and has difficulty in the optimization of deep
structures. To overcome the mentioned issues, deep learning tech-
Ai,t the ith locally constant SDPs nique was developed and applied to an industrial CDU by Shang et
ˆk
A the state vector in polynomial Spline at the kth al. [17] in order to build deep NN-based soft sensors. The predic-
sample tions made by the model match real values better than traditional
At state vector data-driven modeling approaches such as the single hidden layer
ai,t the ith SDP in the regression model neural network, SVM, PLS and NNPLS while the model had a rela-
et white noise tively higher training time.
I identity matrix There are many studies that have been applied modified lin-
K kernel function ear and non-linear approaches to CDU soft sensors, e.g. evolving
Lc concentrated likelihood function fuzzy Takagi–Sugeno models [18,19], modified nonlinear general-
N number of data points in time series ized ridge regression (MNGRR) [20], auto-regressive moving av-
nsi number of state that the ith SDP is a function of erage with exogenous inputs (ARMAX), nonlinear auto-regressive
them model with exogenous inputs (NARX) and Hammerstein-Wiener
Pk covariance matrix of SDP estimation (HW) models [21] and output error (OE) and neuro-fuzzy models
p number of all parameters of the model in At [22].
pi number of parameters in Ai,t The literatures disclose the massive use of different modeling
q local polynomial order methods for estimating quality properties of CDU products. The
q j,i local polynomial order of the ith SDP with respect soft sensor modeling methods for CDU mainly focus on using pres-
to xj,i, t sure, temperature, flow rate, etc. and the study on using multiple
Rk variance of prediction noise normalized by σ 2 temperature measurements have not been taken into account. The
S i,t Spline function of ith SDP temperature as an accessible and measurable variable is widely
t sample time used to estimate product qualities in distillation columns [23–25].
Uk LIV matrix at the kth sample This motivates the design of temperature data-driven soft sensors
Um,k LIV correspondent to the mth regressor at the kth to infer qualities based on the process model and available tem-
sample perature measurements [26].
W local weighting matrix Using varying parameter models such as state dependent pa-
xj,i, t the jth element of ith vector of states rameter (SDP) models as a novel data-driven approach to train soft
y the time series vector sensor models introduced by Gharehbaghi and Sadeghi [27] and
yi the ith output in y Bidar et al. [28]. The modeling approach incorporates process infor-
yt output mation into the model while at the same time provides a good ex-
Zk regressor defined for polynomial Spline planation of data. In this way, SDP-based soft sensors have shown
z vector of regressors successful applications for identification and estimation of indus-
zi,t regressor correspondent to the ith SDP trial processes, where SDP method outperforms other traditional
0 zeros vector data-driven methods like PCR, PLS, ANN, SVR and so on, due to its
 distance between two knots remarkable ability to describe the behavior of non-linear systems.
εk prediction error In SDP estimations, a state must have two special properties to
λj,i bandwidth correspondent to xj,i, t be called an instrumental variable (IV). Each IV should be corre-
σ2 variance of error lated with correspondent regressors as much as possible while it
should be correlated with the estimated error and other regressors
as little as possible. Otherwise, the estimation of each SDP affects
relationships and cannot handle the dynamic of nonlinear pro- the estimation of other SDPs and because they are functions of dif-
cesses. Therefore, Shang et al. [11,12] suggested dynamic PLS-based ferent state variables, this will cause distortion in the final estimate
(DPLS) soft sensor modeling approach with temporal smoothness. of SDP. In this case, there is no need to sort after finding the IVs
They compared different approaches to improve DPLS soft sensors and consequently there is no need for a back-fitting algorithm in
through a CDU. The proposed soft sensors were established to order to eliminate the effects of other states and regressors on the
predict the ASTM temperatures by using process variables such estimation of the desired parameter.
as temperatures, pressures, flows, etc. Moreover, non-linear exten- With regard to SDP modeling method, this paper presents the
sions of PCA models such as kernel PCA (KPCA) was developed by design and implementation of a novel data-driven soft sensor us-
Li et al. [13] for estimation of dry point and flashpoint of avia- ing the technique of instrumental variables (IV) to introduce a new
tion kerosene in the atmospheric distillation column. The results method of SDP estimation termed local instrumental variable (LIV).
showed that the performance of the model was in good agree- An extensive evaluation of the proposed method to soft sensor de-
ment with lab measurements. However it should be noted that the sign is conducted using a simulated industrial crude distillation
KPCA approach suffers from the difficulty in choosing nonlinear unit, and demonstrate its good performance that is comparable to
parameters. the available Aspen soft sensors. The proposed method is used to
The complexity and non-linearity of CDU prohibits the practi- predict 95%ASTM-D86 of side products in atmospheric crude dis-
cal use of linear methods and that motivates researchers to con- tillation unit. The soft sensor model is applied to the Aspen dy-
sider non-linear soft sensing approaches, such ANN. Dam and Saraf namic model of the column using MATLAB-Simulink while the pro-
[14] proposed an ANN-based soft sensor for prediction of ASTM portional integral plus (PIP) control structure is implemented. The
temperatures, specific gravities and Flash Points of CDU prod- Aspen soft sensors represent the values of ASTM-D86 of side prod-
ucts. Liu et al. [15] developed and deployed a neural network- ucts, which are considered as the reference values for validation
based data-driven soft sensor for estimation of ASTM 90% distil- of proposed soft sensor. The prediction results of the SDP-based
lation temperature (D90). Rogina et al. [16] also developed neural soft sensors are compared with Aspen soft sensors and further val-
network-based models such as LNN, MLP and RBF for light naphtha idated the soft sensing model when applying it online over simu-
vapor pressure (RVP) estimation. However, ANN demonstrates sev- lated unit.

Please cite this article as: B. Bidar et al., A data-driven soft-sensor for monitoring ASTM-D86 of CDU side products using local instru-
mental variable (LIV) technique, Journal of the Taiwan Institute of Chemical Engineers (2018), [Link]
JID: JTICE
ARTICLE IN PRESS [m5G;January 23, 2018;13:11]

B. Bidar et al. / Journal of the Taiwan Institute of Chemical Engineers 000 (2018) 1–11 3

2. Methodology description where zt is the new vector of regressors and At is the vector of
the parameters of local polynomials describing SDPs, both at time
Data-based mechanistic (DBM) modeling philosophy as an al- sample t and are defined as,
ternative to conventional modeling approaches was introduced by  
zt = z1,t S1,t z2,t S2,t ··· zn,t Sn,t (5)
Young [29]. In the philosophy, the non-stationary characteristics of
the system are expressed by a varying parameter model with a with definition of Z = [zT1 zT2 zT3 · · · ]T = [Z1 Z2 · · · Z p ],
simple structure, which its parameters are either function of time 
where p = ni=1 pi is the number of regressors in zt or equiv-
or some states of the system. These types of models are called alently number of all parameters of the model in At , and y =
state dependent parameter (SDP) models. Suppose DBM modeling [y1 y2 y3 · · · ]T , the solution to IV estimation of Eq. (4) at the
for a non-linear system suggests an SDP model, which at each time kth sample then can be written as [34,38],
sample t, is as follows:  −1
 
n
ˆ k = UT Z
A k UTk y
yt = ai,t · zi,t + et  −1   −1
i=1   , ∀t (1) Pk = σˆ k UTk Z
2
UTk Uk ZT Uk
ai,t = ai x1,i,t , x2,i,t , . . . , xnsi ,i,t  
σˆ k2 = var y − ZAˆ k (6)
Here yt is the model output,n is the number of SDPs/regressors,
zi,t is the ith regressor and ai (·) is the ith SDP that is a function of ˆ k , at the kth
where Pk is the covariance matrix of SDP estimation, A
nsi correspondent states (xj,i, t , j = 1, 2, …, nsi ). When ai,t is as- sample and Uk as defined by Eq. (7) plays the role of IV matrix at
sumed to be constant and not state dependent, nsi = 0. et = N(0, that sample, it is called Local IV (LIV),
σ 2 ) is a zero mean white Gaussian distributed unknown noise with  
Uk = U1,k U2,k ··· U p,k (7)
variance σ 2 .
Um,k is the LIV correspondent to Zm for each m = 1, 2, …, p. The
ideal IV candidate, in SDP estimation, is an IV, which is not corre-
2.1. Local instrumental variable
lated with non-correspondent regressors. It is because suchUk sets
the off-diagonal elements of UTk Z equal to zero and the resulting
Instrumental variables firstly introduced to define a method,
for identification and estimation of model parameters, that has a UTk Z is a diagonal matrix, so the estimation of each SDP is not af-
simple structure as ordinary least square (OLS) method but can fected by the estimation of other SDPs. Therefore, simultaneous es-
support special requirements in the estimation of model parame- timation of SDPs in their own state variable spaces will be possible
ters, e.g. dynamic transfer function (DTF) modeling [30]. There are without the back-fitting requirement. Following theorem can help
many different methods for estimating of SDPs in Eq. (1) ([31–33] to find the proper IVs with this specification.
and much more) including nonparametric and parametric meth- Theorem. Having local weighting matrix, Wm,k , suppose that
ods. Some of these methods are generalized random walk (GRW: ZT Wm,k Z is not singular. Defined by Eq. (8), Um,k is a good LIV can-
[34]) using Kalman filter (KF) concepts, functional coefficient (FC) didate for the SDP estimation.
modeling [35] using local polynomial modeling (LPM: [36]) con-  −1
cepts, Spline methods [37], local models (LM), etc. Polynomial con- Um,k = Wm,k Z ZT Wm,k Z Im m = 1, 2, . . . , p (8)
cepts are almost used in all of these methods. The nonparametric
Im is the mth column of identity matrix, I, of dimension p.
methods suffering from using the back-fitting algorithm, because
the functionality of each SDP affects the estimation of other SDPs, Proof. To prove, it must be shown that with introduced LIV,UTk Z
that produce other problems associated with the back-fitting al- or equivalently ZT Uk is diagonal. Pre-multiplication of Eqs. (7) and
gorithm. In the current study, the polynomial modeling concepts (8) by ZT , yields Eqs. (9) and (10) respectively.
combined with IV concepts have been used to introduce an esti-  
mation method of SDPs such that estimating the SDPs, simultane- ZT Uk = ZT U1,k ZT U2,k ··· ZT U p,k (9)
ously.
If it is assumed that the functionality of each SDP from its  −1
ZT Um,k = ZT Wm,k Z ZT Wm,k Z Im = Im m = 1, 2, . . . , p (10)
correspondent states can be defined by a local polynomial in the
state variable space, then there is a chance to locally estimate Substitution from Eq. (10) into Eq. (9) gives ZT Uk =
the parameters of these polynomials by the implementation of IV [I1 I2 · · · I p ] = I that is identity matrix and definitely
method. By this assumption, ai,t can be defined as, diagonal. Using the proposed LIV matrix, the estimated SDPs in
ai,t = Si,t Ai,t i = 1, 2, . . . , n (2) Eq. (6) then can be formulated as simple as,
ˆ k = UT y
A
where Ai,t is the vector of locally constant parameters of the poly- k  (11)
nomial describing ai,t and Si,t is defined by Eq. (3): Pk = σˆ k2 UTk Uk

Si,t = S0,i,t  S1,i,t  S2,i,t  · · ·  Snsi ,i,t , i = 1, 2, . . . , n The local weighting matrix, Wm,k , can be defined generally to
 q j,i  face some requirements in the model. A simple definition of the
S j,i,t = 1 x j,i,t x2j,i,t ··· x j,i,t , j = 0, 1, 2, . . . , nsi
local weighting matrix in SDP estimation can be a diagonal matrix,
q0,i = 0 (3) which its diagonal elements are the values of kernel function cor-
respondent to the ith SDP at the kth sample. e.g. Wm,k is defined
where  is the Kronecker tensor product symbol and qj,i is the or- as,
der of polynomial describing ai,t with respect to xj,i, t . The number
nsi
⎧    
of parameters in Ai,t will be obtained by pi = j=0 (1 + q j,i ). All ⎨Wm,k = diag K i,k,t ∀t
member of Ai,t are state dependent functions of xj,i, t , j = 1, 2, …, 
nsi
x j,i,t −x j,i,k
2 i
−1 
i (12)
⎩i,k,t = λ j,i 0< pv < m ≤ pv
nsi . Substitution from Eq. (2) yields the model (Eq. (1)) in vector j=1 v=1 v=1
form as,
K(·) is the kernel function and λj,i is the bandwidth correspon-
yt = zt At + et (4) dent to xj,i, t . λj,i acts as hyper-parameter that must be found in

Please cite this article as: B. Bidar et al., A data-driven soft-sensor for monitoring ASTM-D86 of CDU side products using local instru-
mental variable (LIV) technique, Journal of the Taiwan Institute of Chemical Engineers (2018), [Link]
JID: JTICE
ARTICLE IN PRESS [m5G;January 23, 2018;13:11]

4 B. Bidar et al. / Journal of the Taiwan Institute of Chemical Engineers 000 (2018) 1–11

optimization procedure. Substituting of estimated Aˆ k in Eq. (2), we

obtain SDP estimation and its covariance at the kth sample as,

ˆ
aˆ k = SkA
(13)
cov aˆ k = Sk PSTk
where aˆ k and Sk are defined as Eq. (14) with 0 as appropriate vec-
tor of zeros as required,
 T
aˆ k = aˆ1,k aˆ2,k ··· aˆn,k
⎡ ⎤
S1,k 0 ··· 0
⎢ 0 S2,k ··· 0 ⎥ (14)
Sk = ⎢ . .. ⎥
⎣ .. ..
.
..
. .
⎦.
0 0 ··· Sn,k

2.2. Hyper-parameter optimization

Fine tuning of the hyper-parameters (e.g., λj,i ) plays an impor-

tant role to minimize the errors and achieve an accurate model.
Different optimization criteria can be applied to SDP models to
find appropriate hyper-parameters. Here we focus on one of the
most popular criteria; maximum likelihood (ML) approaches. The
ML optimization approach is accomplished based on prediction er-
ror decomposition (PED), which yields to one-step-ahead predic-
tion errors. If εˆk is the residual at the kth sample when that sam-
ple is removed from calculation and Rk is the covariance of εˆk , then
the concentrated likelihood can be expressed as,
  
1 1  εˆk2
N N
1
log(Lc ) = − log (Rk ) + log (15)
2 N N Rk
k=1 k=1
Fig. 1. Process flow sheet of atmospheric distillation column (ADU) [39].
which should be minimized respect to the unknown hyper-
parameters in order to obtain their ML estimates. Since Likelihood
Table 1
function is a nonlinear function of the hyper-parameters, the mini- Input variables of PIP controller and their steady-state values for operating capacity
mization needs to be carried out numerically. This is accomplished 60,0 0 0 bbl/day.
by initiating the optimization with hyper-parameter estimates by
Variable Description Initial value
either the user or set to some default values.
u1 Side-draw of naphtha 29748.8 lb/h
3. Case study: crude distillation unit u2 Side-draw of kerosene 128254.6 lb/h
u3 Side-draw of gas oil 122868.1 lb/h
u4 Furnace duty 110,173,828 Btu/h
3.1. Case description u5 Reflux split fraction 0.61676
u6 Heat removed in pumparound1 −22,428,847 Btu/h
The crude distillation unit usually consists of an atmospheric u7 Heat removed in pumparound2 −26,172,804 Btu/h
(ADU) and vacuum distillation units (VDU). An ADU is one of the
critical unit operations for the petroleum industry. In this study,
Table 2
an industrial atmospheric distillation column of Shiraz refinery and Output variables.
relevant parameters in [39] are considered as a case study in or-
Variable Description Set point (°F)
der to generate simulated data. The column utilizes Aspen dy-
namic and MATLAB-Simulink to link Aspen simulation with PIP T1 Overhead temperature 295.8
control configuration. Given the PIP control structure to maintain y1 95%ASTM-D86 of naphtha 385.7
y2 95%ASTM-D86 of kerosene 502.3
column product qualities at some prescribed set points is a moti- y3 95%ASTM-D86 of gas oil 623
vation of soft sensor design. Fig. 1 shows the process flow sheet
of the ADU, which consists of a crude tower, a top condenser, two
mid-tower condensers (pumparounds) and three side strippers. It
includes 43 trays and operates at the maximum operating capac- a standard approach to analyze the product quality, which is a
ity 60,0 0 0 bbl/day. The main products of the column are naphtha, cumber-some and time-consuming procedure. The 5%ASTM-D86 is
kerosene and gas oil, which are taken from three side strippers. a quality index for light fractions, which are further maintained
Based on the degree of freedom (see Appendix A in [39]), there by steam rate in side strippers within specifications. Thereby, the
are 7 manipulated (u1 to u7 ) and 4 controlled variables (Tables 1 important quality index of “95%ASTM-D86”, limits the amount of
and 2). The 95%ASTM-D86 of naphtha (y1 ), kerosene (y2 ) and gas heavy product that can be extracted in a side products. In this
oil (y3 ) are output variables, which are supposed to be determined study, 95%ASTM-D86 of naphtha, kerosene and gas oil are selected
by the soft sensor (see Fig. 5). as the quality variables for soft sensor development.
The ultimate objective is to find appropriate soft sensing model
3.2. Setting for soft sensing modeling between the measured tray temperatures inside the column and
the quality of the side products. The satisfactory performance of
3.2.1. Variable selection soft sensors is likely to be achieved if only input variables that are
The main goal of controlling ADU is to maintain the product most sensitive to the output variables are employed. As the tray
quality within the specified range. The ASTM-D86 temperature is temperatures are correlated, the independent few measurements

Please cite this article as: B. Bidar et al., A data-driven soft-sensor for monitoring ASTM-D86 of CDU side products using local instru-
mental variable (LIV) technique, Journal of the Taiwan Institute of Chemical Engineers (2018), [Link]
JID: JTICE
ARTICLE IN PRESS [m5G;January 23, 2018;13:11]

B. Bidar et al. / Journal of the Taiwan Institute of Chemical Engineers 000 (2018) 1–11 5

at appropriate locations should be selected to provide fairly ac- are considered as states in one parameter and the corresponding
curate information about quality and avoid the over-fitting prob- regressor is chosen as one. Although the ADU is a multi-output
lem [40]. Correlation analysis with backward elimination method process, it is treated as three single-output processes, so thaty1 ,
is carried out to remove redundant tray temperatures. Correlation y2 and y3 are modeled, separately. The training procedure is re-
matrix using Pearson correlation coefficients is calculated from the peated for each modeling process as long as the performance error
available training data. Based on correlation color map, more rel- does not further decrease. Finally, the best model is selected based
evant input variables are set into several groups, where variables on least error.
within the group are highly correlated, but variables between the
groups are weakly correlated. This gives a visual way of classifying 3.3. Soft sensor performance evaluation
large data set of variables into several uncorrelated groups [41].
The high correlated input variable with the output, which is the To assess the model performance of the soft sensors, three per-
most important variable from a process point of view, is selected formance indexes are used, namely root mean square error (RMSE),
as a representative variable of each group. The soft sensing model mean absolute error (MAE), the coefficient of determination (R2 )
is trained using all selected variables and in the following, vari- and adjusted R2 as defined:
ables are subsequently eliminated one at a time from the model 
using backward elimination method. If the model performance de- 1 2
RMSE = yi − yˆi (16)
teriorates, the eliminated variable is considered important and re- N
introduced in the model. N  
i=1
yi − yˆi 
MAE = (17)
3.2.2. Data collection N
Since there is no possibility of making changes in the industrial N  2
unit to collect data, Aspen simulated model was accepted as a sub- 2 i=1 yˆi − ȳ
R =1− N (18)
stitute for the real industrial unit. However, the results of steady- (yi − ȳ )2
i=1
state simulation for 95%ASTM-D86 with laboratory data of Shiraz   
refinery have been validated with a slight error (with a maximum 1 − R2 ( N − 1 )
error of 2%) in [39]. In order to make the soft sensing model have a R2adj =1− (19)
(N − k − 1 )
good ability of generalization, the simulated data are obtained by
linking the MATLAB-Simulink to Aspen dynamic model. The data where N is the number of data samples, k is the number of inde-
collection for estimating is required to stimulate the system with pendent regressors or predictors and yi ,yˆi , ȳ and yˆ¯ are referred to
a random signal around specific operating conditions. For this pur- as the real value, predicted value, mean values of yand yˆ , respec-
pose, the Monte Carlo method is applied to generate required sim- tively.
ulated data. The set point values of 95%ASTM-D86 for naphtha,
kerosene and gas oil are shown in Table 2. For data generation, 4. Results and discussion
the set points of product qualities are randomly sampled (in the
range of ±20 °F), which have been used to perturb the system and Suppose Eq. (1) for each product quality (95%ASTM-D68Naphtha ,
provide input–output pairs suitable for the identification of soft 95%ASTM-D86Kerosene , 95%ASTM-D86Gas oil ) can be expressed in fol-
sensor model. In order to keep the system in the vicinity of such lowing forms:
a selected operating condition, the PIP controller has been em-
ployed. Then, for each group samples, the simulation is performed y1,t = a1,t {T1 , T4 , T14 , T20 , T28 , T32 , T37 , T38 } × 1 + et (20)
as long as steady-state operating condition (ASTM standard devi-
y2,t = a2,t {T2 , T3 , T19 , T20 , T21 , T32 , T36 , T43 } × 1 + et (21)
ation lower than 1 °F) achieved to obtain the corresponding tem-
perature profile. It assumes that there are no changes in pressure
y3,t = a3,t {T2 , T3 , T31 , T32 , T34 , T43 } × 1 + et . (22)
and tray efficiencies of the column. A total of 10 0 0 steady-state
random set of data samples have been collected, which is further After the initial assessment of input variables, the most rele-
divided into the training dataset (500 samples) and testing dataset vant tray temperatures are selected based on proposed strategy
(500 samples). described in Fig. 2. The training procedure is repeated 11 times
for both naphtha and kerosene models and 10 times for gas oil
3.2.3. Selection of model structure and regression parameters model. Finally, model configurations with the best performance
The regression modeling between tray temperatures and each are selected for three product quality, which is achieved by op-
95%ASTM-D86 temperature for naphtha (y1 ), kerosene (y2 ) and gas timized bandwidths. The optimized bandwidths corresponding to
oil (y3 ) is performed through Eq. (1). The soft sensing models for each state are obtained using the local polynomial with the first
estimation of each quality variables in side stream products are order for all selected states. Detailed results and performance in-
built individually. The variable selection steps are shown using a dexes of proposed soft sensors are given in Tables 5 and 6. It is
flowchart in Fig. 2. All tray temperatures of the column (T1 –T43 ) observed from Table 5 that the number of input variables is de-
are initially selected as the input of soft sensing procedure. Fig. 3 creased as compared to the initially selected temperatures. The
shows the correlation color map of 43 temperature variables, RMSE index implies that enough information is provided to soft
where temperatures are grouped by correlation. The correlation sensors thus leading to accurate models. Also, the MAE for each
coefficients between inputs (tray temperature variables) and out- soft sensor models is less than 0.5 °F, which exhibit very good co-
puts (95%ASTM-D86 temperature variables) are listed in Table 3. incidence between the model prediction and simulated data. The
The map clearly identifies eight groups, reported in Table 4. R2 values are close to 1, which means that it is a perfect predic-
There is eight representative temperatures for naphtha tion model. Moreover, the adjusted R2 values are similar to the R2
and kerosene and six representative temperatures for gas oil values, which confirm that non-significant variables have not been
(see Table 4), which are selected based on the maximum correla- added to the model, because in that case the gap in R2 and ad-
tion coefficient from each group (see highlighted cells in Table 3). justed R2 increased.
Therefore, a different set of input variables are considered for Graphical overviews are presented in Fig. 4 showing good
each soft sensing model. Moreover, all representative temperatures agreement with the mentioned performance indexes. The robust

Please cite this article as: B. Bidar et al., A data-driven soft-sensor for monitoring ASTM-D86 of CDU side products using local instru-
mental variable (LIV) technique, Journal of the Taiwan Institute of Chemical Engineers (2018), [Link]
JID: JTICE
ARTICLE IN PRESS [m5G;January 23, 2018;13:11]

6 B. Bidar et al. / Journal of the Taiwan Institute of Chemical Engineers 000 (2018) 1–11

Fig. 2. Schematic diagram of the soft sensor training based on SDP-LIV method and correlated variables.

modeling performance of SDP-based soft sensing models further on the SDP-based soft sensor outputs. The optimum values of
discloses their potential for the practical online implementation weighting matrices (Q and R) are obtained based on minimizing of
under accurate and stable predictions. Fig. 5 shows the location of the ITAE (for detailed information see Ref. [39]). The scenario for
the soft sensor in PIP control loop and its application to the col- calculating ITAE is a +2% step change in feed rate, which starts by
umn. MATLAB-Function blocks are utilized to implement soft sen- initial values of weighting matrices and continues by changing the
sors into Aspen dynamic model. Q and R based on Nelder–Mead simplex algorithm until the min-
The training procedure is done with training data set and the imum value of ITAE is reached. The optimized Q and R are given
static models of soft sensors are obtained for each product qual- by the following matrices using proposed soft sensors:
ities. The resultant models then used to predict 95%ASTM-D86 of
side products in an online adaptation over simulated model with Q = diag(19.07 × 10−10 , 3.5 × 10−10 , 1.11 × 10−10 , 6.15
PIP control system. The PIP controller parameters are reset based ×10−10 , 1, 1, 1, 1 ) (23)

Please cite this article as: B. Bidar et al., A data-driven soft-sensor for monitoring ASTM-D86 of CDU side products using local instru-
mental variable (LIV) technique, Journal of the Taiwan Institute of Chemical Engineers (2018), [Link]
JID: JTICE
ARTICLE IN PRESS [m5G;January 23, 2018;13:11]

B. Bidar et al. / Journal of the Taiwan Institute of Chemical Engineers 000 (2018) 1–11 7

Fig. 3. Correlation color map for grouping temperature variables based on correlation matrix (higher color density, higher correlation).

Table 3
Correlation coefficients between 95%ASTM-D86 temperatures for naphtha, kerosene and gas oil with tray temperatures.

T1 T2 T3 T4 T5 T6 T7 T8 T9 T10 T11

y1 −0.761 −0.647 −0.011 0.740 0.895 0.930 0.943 0.949 0.953 0.956 0.959
y2 0.050 0.252 0.450 0.292 0.146 0.077 0.038 0.014 −0.002 −0.013 −0.021
y3 −0.105 −0.209 −0.303 −0.170 −0.072 −0.027 −0.003 0.011 0.021 0.027 0.031
T12 T13 T14 T15 T16 T17 T18 T19 T20 T21 T22
y1 0.962 0.964 0.965 0.965 0.955 0.954 0.948 0.906 0.309 −0.258 −0.426
y2 −0.027 −0.030 −0.030 −0.013 0.079 0.085 0.124 0.287 0.919 0.959 0.901
y3 0.034 0.036 0.037 0.037 0.032 0.031 0.028 0.015 −0.046 −0.073 −0.079
T23 T24 T25 T26 T27 T28 T29 T30 T31 T32 T33
y1 −0.481 −0.503 −0.516 −0.528 −0.546 −0.580 −0.544 −0.542 −0.557 −0.462 −0.311
y2 0.874 0.861 0.854 0.845 0.833 0.802 0.824 0.815 0.749 0.369 0.014
y3 −0.080 −0.081 −0.079 −0.072 −0.049 0.029 0.054 0.113 0.283 0.754 0.916
T34 T35 T36 T37 T38 T39 T40 T41 T42 T43
y1 −0.250 −0.246 −0.271 −0.336 −0.707 −0.689 −0.679 −0.668 −0.648 −0.624
y2 −0.135 −0.188 −0.203 −0.196 −0.088 −0.095 −0.098 −0.103 −0.110 −0.115
y3 0.933 0.928 0.917 0.892 0.628 0.647 0.658 0.670 0.690 0.711

Table 4
Grouping variables based on correlation matrix.

Variable no. Group Selected variable for naphthaa Selected variable for kerosenea Selected variable for gas oila

T1 , T2 1 T1 T2 T2
T3 , T4 2 T4 T3 T3
T5 –T19 3 T14 T19 –
T20 4 T20 T20 –
T21 –T31 5 T28 T21 T31
T32 6 T32 T32 T32
T33 –T37 7 T37 T36 T34
T38 –T43 8 T38 T43 T43
a
Representative variable selected from each group for using in the variable selection procedure.

Table 5
Selected states and corresponding optimal bandwidths of SDP-based soft sensors.

Product quality Selected states Optimized bandwidths for each state

95%ASTM-D86 of naphtha T4 , T14, T28 8.4442 × 10−2 8.1877 × 10−3 5.4174

95%ASTM-D86 of kerosene T20 , T21 , T43 5.0288 × 10−2 6.3121 × 10−3 8.3991 × 10−1
95%ASTM-D86 of gas oil T32 , T34 , T43 9.0873 1.5965 × 10−2 2.8875 × 10−2

Please cite this article as: B. Bidar et al., A data-driven soft-sensor for monitoring ASTM-D86 of CDU side products using local instru-
mental variable (LIV) technique, Journal of the Taiwan Institute of Chemical Engineers (2018), [Link]
JID: JTICE
ARTICLE IN PRESS [m5G;January 23, 2018;13:11]

8 B. Bidar et al. / Journal of the Taiwan Institute of Chemical Engineers 000 (2018) 1–11

Fig. 4. Quality prediction results by SDP-based soft sensors on testing data set (a) 95%ASTM-D86 of naphtha, (b) 95%ASTM-D86 of kerosene and (c) 95%ASTM-D86 of gas oil.

Fig. 5. ASTM-D86 soft sensors and PIP controller in MATLAB-Simulink.

Please cite this article as: B. Bidar et al., A data-driven soft-sensor for monitoring ASTM-D86 of CDU side products using local instru-
mental variable (LIV) technique, Journal of the Taiwan Institute of Chemical Engineers (2018), [Link]
JID: JTICE
ARTICLE IN PRESS [m5G;January 23, 2018;13:11]

B. Bidar et al. / Journal of the Taiwan Institute of Chemical Engineers 000 (2018) 1–11 9

Table 6
Performance indexes of three soft sensors on training and testing data sets.

SDP- based Soft sensor Training performance indexes Testing performance indexes

R2 R2adj RMSE MAE R2 R2adj RMSE MAE

Naphtha 95%ASTM-D86 1 0.99999 0.0106 0.0051 0.9989 0.99889 0.1781 0.1361

Kerosene 95%ASTM-D86 1 0.99999 0.0167 0.0054 0.9999 0.99989 0.3156 0.2536
Gas oil 95%ASTM-D86 1 0.99999 0.0676 0.0396 0.9784 0.97836 0.6147 0.4851

Fig. 6. Output performance for Aspen soft sensors and SDP-based soft sensors (a) +2% step change in feed rate (b) −2% step change in feed rate.

would be reached to the soft sensor predicted values after the sys-
tem has achieved to its steady-state condition.
R = diag(7.17 × 10 −4
, 4.5 × 10 −5
, 5.42 × 10 −5
, 3.51
In this case, after about 2 h, the 95%ASTM-D86 temperatures,
×10 −11 7
, 9.85 × 10 , 8.29 × 10 −9
, 1.64 × 10−8 ) (24) which are predicted by SDP-based soft sensors reached to the sim-
In order to evaluate the performance of the SDP-based soft ulated values, that shows faster response than Aspen soft sensors
sensor, the column is disturbed by a step change in feed rate at and therefore provides better disturbance rejection. According to
t = 0.5 h while the set points of PIP controller were in the set the output performances (Fig. 6a and b), the 95%ASTM-D86 tem-
point values. Fig. 6 provides comparative performances of Aspen peratures have been suddenly changed as well as the prediction
and SDP-based soft sensors, which are assessed by subjecting the results of Aspen and SDP-based soft sensors exhibited a time dif-
column to ±2%, feed rate changes from steady-state operating con- ference. However, the settling times of the 95%ASTM-D86 for three
ditions (60,0 0 0 bbl/day). It can be concluded that the 95%ASTM- side products in SDP-based soft sensors model are around 1.5 h
D86 of all side products are reasonably well predicted by SDP- and distortions are removed from the system. The difference in
based soft sensors, and are consistent with prediction results from output performances lie for the following reasons: (a) the sensi-
Aspen soft sensors in the new operating condition. tivity of proposed soft sensor to change the temperature profile;
In online mode, the temperatures were sensed and transferred (b) the soft sensing model is built on static data while the perfor-
to soft sensing models (It is about a few seconds); the measured mance has been tested in a dynamic condition; The overshoot in
outputs from soft sensors were also taken in a few seconds. Al- the ASTM temperatures for naphtha, kerosene and gas oil is rel-
though, the soft sensing models are static, Fig. 6 demonstrates that atively small and the damping is very fast as shown in Fig. 6a
they have been able to operate well in online mode and the out- and b. The overshoot for positive step change is about 2 °F, 5.5 °F
put qualities can be reached to the expected values in less than and 10.8 °F and for negative step change about 4.5 °F, 6.8 °F and
3 h. However, in the laboratory analysis, from sampling to report 7.7 °F for ASTM-D86Naphtha , ASTM-D86Kerosene and ASTM-D86Gas oil ,
the ASTM values lasted for at least 8 h. respectively. The corresponding percent errors of ASTM prediction
As the model training is done offline with steady-state data, and by proposed soft sensors, compared with the Aspen results (simu-
the embedded soft sensor runs using built-in steady-state model, lated values) are listed in Table 7.
the prediction results in each sample time represent the final value The percentages are very close to zero that means the match
of 95%ASTM-D86 according to the tray temperature values in the would be perfect; an indication that the discrepancy between pro-
column. In the other words, if the temperatures through the col- posed soft sensor and Aspen results would be absent. Moreover,
umn are fixed at their corresponding values, the product qualities compared to PIP controller results with Aspen soft sensors, the PIP

Please cite this article as: B. Bidar et al., A data-driven soft-sensor for monitoring ASTM-D86 of CDU side products using local instru-
mental variable (LIV) technique, Journal of the Taiwan Institute of Chemical Engineers (2018), [Link]
JID: JTICE
ARTICLE IN PRESS [m5G;January 23, 2018;13:11]

10 B. Bidar et al. / Journal of the Taiwan Institute of Chemical Engineers 000 (2018) 1–11

Table 7 References
The difference between Aspen and SDP-based soft sensor results after 5 h.

+2% step change in −2% step change in [1] Luyben WL. Distillation design and control using Aspen simulation. New Jer-
sey: John Wiley & Sons; 2006.
Quality variable feed rate feed rate
[2] Dave DJ, Dabhiya MZ, Satyadev SVK, Ganguly S, Saraf DN. Online tuning of a
95%ASTM-D86 of naphtha 0.01% 0.01% steady state crude distillation unit model for real time applications. J Process
95%ASTM-D86 of kerosene 0.003% 0.013% Control 2003;13:267–82. [Link]
95%ASTM-D86 of gas oil 0.005% 0.003% [3] Ujević Ž, Mohler I, Bolf N. Soft sensors for splitter product property estima-
tion in CDU. Chem Eng Commun 2011;198:1566–78. [Link]
00986445.2011.556692.
[4] Galicia HJ, He QP, Wang J. A reduced order soft sensor approach and its ap-
plication to a continuous digester. J Process Control 2011;21:489–500. http:
controller has acceptable output performances for a step change in //[Link]/10.1016/[Link].2011.02.001.
the feed rate and keeps the output variables at the desired values [5] Liu J. Developing a soft sensor based on sparse partial least squares with vari-
by using proposed soft sensors. The results show that SDP-based able selection. J Process Control 2014;24:1046–56. [Link]
jprocont.2014.05.014.
soft sensors successfully model the relationship between the tem- [6] Fortuna L, Graziani S, Xibilia MG. Soft sensors for product quality monitoring
peratures inside the column with ASTM-D86 temperatures of prod- in debutanizer distillation columns. Control Eng Pract 2005;13:499–508. http:
ucts and can predict the ASTM index much faster than what is ac- //[Link]/10.1016/[Link].2004.04.013.
[7] Gonzaga JCB, Meleiro LAC, Kiang C, Filho RM. ANN-based soft-sensor for real-
tually done by laboratory analysis.
time process monitoring and control of an industrial polymerization process.
Comput Chem Eng 2009;33:43–9. [Link]
2008.05.019.
5. Conclusion [8] Jain P, Rahman I, Kulkarni BD. Development of a soft sensor for a batch distil-
lation column using support vector regression techniques. Chem Eng Res Des
In this work, a novel soft sensor modeling approach based on 2007;85:283–7. [Link]
[9] Liu Y, Chen J. Integrated soft sensor using just-in-time support vector regres-
SDP models and the technique of instrumental variables is pro- sion and probabilistic analysis for quality prediction of multi-grade processes. J
posed to deal with the issue of the back-fitting algorithm in SDP Process Control 2013;23:793–804. [Link]
models. Since, in this method, the effect of other regressors and 008.
[10] Wang D, Liu J, Srinivasan R. Data-driven soft sensor approach for quality pre-
states on the estimation of a parameter is eliminated, there is no diction in a refining process. IEEE Trans Ind Inform 2010;6:11–17. [Link]
need to sort and consequently no need to the back-fitting algo- [Link]/10.1109/TII.2009.2025124.
rithm. Furthermore, as the LIV method can be used for interpola- [11] Shang C, Gao X, Yang F, Lyu W, Huang D. A comparative study on improved
DPLS soft sensor models applied to a crude distillation unit. Proceedings of
tion directly without parameterization, it can handle estimation of the 9th IFAC symposium on advanced control of chemical processes; ADCHEM
SDPs and also used in full nonparametric state dependent trans- 2015; Whistler, Canada: Elsevier. [Link]
fer function (SDTF) modeling with the associated improved con- [12] Shang C, Huang X, Suykens JAK, Huang D. Enhancing dynamic soft sensors
based on DPLS: a temporal smoothness regularization approach. J Process Con-
trol models. The proposed method is applied to estimation of the
trol 2015;28:17–26. [Link]
95%ASTM-D86 of three side products (naphtha, kerosene and gas [13] Li Q, Du Q, Ba W, Shao C. Multiple-input multiple-output soft sensors based on
oil) in an industrial atmospheric crude distillation column. Corre- KPCA and MKLS-SVM for quality prediction in atmospheric distillation column.
Int J Innov Comput I 2012;8:8215–30.
lation analysis with backward elimination method is carried out to
[14] Dam M, Saraf DN. Design of neural networks using genetic algorithm for on-
select best input variables. Three SDP-based soft sensors are sepa- line property estimation of crude fractionator products. Comput Chem Eng
rately designed in an offline manner for each product quality with 2006;30:722–9. [Link] 05.12.0 01.
steady-state data of the column. The results obtained under the [15] Liu J, Srinivasan R, SelvaGuru PN. Practical challenges in developing
data-driven soft sensors for quality prediction. Comput Aided Chem Eng
MATLAB/Simulink integration based on directly measured temper- 2008;25:961–6. [Link] 7946(08)80166- 6.
atures from the column. The performance of proposed soft sen- [16] Rogina A, Šiško I, Mohler I, Ujević Ž, Bolf N. Soft Sensor for continuous
sors is evaluated with testing data and also by online implementa- product quality wstimation (in crude distillation unit). Chem Eng Res Des
2011;89:2070–7. [Link]
tion in PIP control system over simulated model of unit in Aspen. [17] Shang C, Yang F, Huang D, Lyu W. Data-driven soft sensor development based
The testing results demonstrate that the new technique can ex- on deep learning technique. J Process Control 2014;24:223–33. [Link]
tract nonlinear and complex relationship of process variables with org/10.1016/[Link].2014.01.012.
[18] Macias JJ, Angelov P, Zhou X, editors. A method for predicting quality of the
acceptable accuracy. Moreover, the performance tests of proposed crude oil distillation. Proceedings of the international symposium on evolving
soft sensors for step changes in feed rate are utilized. The results fuzzy systems; 2006: IEEE. [Link]
show that proposed soft sensors can predict the product qualities [19] Macias-Hernandez JJ, Angelov P, Zhou X, editors. Soft sensor for predicting
crude oil distillation side streams using Takagi Sugeno evolving fuzzy mod-
much faster (about 2 h) than Aspen soft sensors. Therefore, it can
els Proceedings of the 2nd annual symposium on nature inspired smart adap-
be concluded that the SDP models with LIV estimation method is tive systems; 2006; Tenerife, Spain: IEEE. [Link]
an efficient one from previously offered models, which can be used 4413939.
[20] Yan X. Modified nonlinear generalized ridge regression and its application to
for simultaneous prediction of ASTM-D86 in an industrial ADU and
develop naphtha cut point soft sensor. Comput Chem Eng 2008;32:608–21.
also used online in process control system. The designed soft sen- [Link]
sors provide a real time estimate of the relevant quality variables, [21] Andrijić ŽU, Bolf N. Soft sensors application for crude distillation unit prod-
thus overcoming the problem of the delay introduced by the lab- uct quality estimation. Goriv Maziva 2011;50:201–14. [Link]
72468.
oratory analysis of ASTM-D86 (about 8 h). It can easily be fur- [22] Novak M, Mohler I, Golob M, Andrijić ŽU, Bolf N. Continuous estimation of
ther implemented as an inferential monitoring of ASTM-D86 in- kerosene cold filter plugging point using soft sensors. Fuel Process Technol
dex, where online ASTM analyzing is not able to perform. This im- 2013;113:8–19. [Link]
[23] Mejdell T, Skogestad S. Estimation of distillation compositions from multi-
provement is achieved through a reasonable decrease in the anal- ple temperature measurements using partial-least-squares regression. Ind Eng
ysis time and procedure cost. Chem Res 1991;30:2543–55. [Link] 0 060a0 07.
[24] Verhaegen M. Identification of the temperature-product quality relationship in
a multi-component distillation column. Chem Eng Commun 1998;163:111–32.
Acknowledgment [Link]
[25] Luyben WL. Evaluation of criteria for selecting temperature control trays in
distillation columns. J Process Control 2006;16:115–34. [Link]
Technical support from Shiraz Refinery, Shiraz, Iran, is grate- 1016/[Link].20 05.05.0 04.
fully acknowledged. [26] Stichlmair JG, Fair JR. Distillation: principles and practices. New York: John Wi-
ley & Sons; 1998.
Funding: This research did not receive any specific grant from
[27] Gharehbaghi H, Sadeghi J. A novel approach for prediction of industrial cat-
funding agencies in the public, commercial, or not-for-profit sec- alyst deactivation using soft sensor modeling. Catalysts 2016;6:93–109. http:
tors. //[Link]/10.3390/catal6070093.

Please cite this article as: B. Bidar et al., A data-driven soft-sensor for monitoring ASTM-D86 of CDU side products using local instru-
mental variable (LIV) technique, Journal of the Taiwan Institute of Chemical Engineers (2018), [Link]
JID: JTICE
ARTICLE IN PRESS [m5G;January 23, 2018;13:11]

B. Bidar et al. / Journal of the Taiwan Institute of Chemical Engineers 000 (2018) 1–11 11

[28] Bidar B, Sadeghi J, Shahraki F, Khalilipour MM. Data-driven soft sensor ap- [35] Fan J, Yao Q. Nonlinear time series: nonparametric and parametric methods.
proach for online quality prediction using state dependent parameter models. New York: Springer; 2003.
Chemom Intell Lab Syst 2017;162:130–41. [Link] [36] Fan J, Gijbels I. Local polynomial modelling and its applications. London: Chap-
2017.01.004. man and Hall; 1996.
[29] Young PC. Data-based mechanistic modeling of engineering systems. J Vib Con- [37] Hastie TJ, Tibshirani RJ. Generalized additive models. London: Chapman and
trol 1998;4:5–28. [Link] 040 0102. Hall; 1990.
[30] Young PC. The instrumental variable method: a practical approach to identifi- [38] Norton JP. An introduction to identification. London: Academic Press; 1986.
cation and system parameter estimation. Barker HA, Young PC, editors, Oxford: [39] Khalilipour MM, Sadeghi J, Shahraki F, Razzaghi K. Nonsquare multivariable
Pergamon Press; 1985. editors. non-minimal statespace-proportional integral plus (NMSS-PIP) control for at-
[31] Young PC. In: Rao TS, editor. Time variable and state dependent parameter mospheric crude oil distillation column. Chem Eng Res Des 2016;113:140–50.
modeling of nonstationary and nonlinear time series. Developments in Time [Link]
Series Analysis. London: Chapman and Hall; 1993. editor p. 374–413. [40] Kano M, Miyazaki K, Hasebe S, Hashimoto I. Inferential control system of dis-
[32] Hastie T, Tibshirani R. Varying-coefficient models. J R Stat Soc Ser B tillation compositions using dynamic partial least squares regression. J Process
1993;55:757–96. [Link] Control 20 0 0;10:157–66. [Link] 0 027-X.
[33] Fan J, Zhang W. Statistical estimation in varying coefficient models. Ann Stat [41] Rahman MM, Imtiaz SA, Hawboldt K. A hybrid input variable selection method
1999;27:1491–518. [Link] for building soft sensor from correlated process variables. Chemom Intell Lab
[34] Young PC. Recursive estimation and time series analysis. Berlin: Springer-Ver- Syst 2016;157:67–77. [Link]
lag; 1984.

Please cite this article as: B. Bidar et al., A data-driven soft-sensor for monitoring ASTM-D86 of CDU side products using local instru-
mental variable (LIV) technique, Journal of the Taiwan Institute of Chemical Engineers (2018), [Link]