Identification of Refinery Distillation Column First

Principles Simulation Models

2023 16th International Conference Management of large-scale system development (MLSD) | 979-8-3503-3790-7/23/$31.00 ©2023 IEEE | DOI: 10.1109/MLSD58227.2023.10303969

Oleg Dranko Elena Baryshnikova Igor Slastenov

V. A. Trapeznikov Institute Institute of Precision Mechanics and Control Institute of Precision Mechanics and Control
of Control Sciences of RAS Saratov Federal Scientific Centre of RAS Saratov Federal Scientific Centre of RAS
Moscow, Russia Saratov, Russia Saratov, Russia
[email protected] [email protected] [email protected]

Abstract—We have verified the principal possibility of identify- Unlike EM, FP simulation models (SM) of TP are built
ing a first principles simulation model of a deethanizer distillation based on knowledge about the structure and principles of the
column of an oil refinery’s typical gas fractionating unit. The object’s functioning. The mentioned features of the models
article analyzes the results of a series of numerical experiments
on identifying this model using an extended Kalman filter. allows to accurately reproduce the object’s behavior in various
Index Terms—identification, extended Kalman filter, simu- process conditions. Nevertheless, the FP SM were barely used
lation modeling, oil refining process, first principles model, in the TP control loop until recent years due to the limitations
distillation column in the calculation speed of such models and the absence of
unified approaches to their identification. However, the devel-
I. I NTRODUCTION opment of IT and simulation tools expands the high-precision
TP SM application range. In this article, we consider the
The last decade has been marked by a noticeable devel- results of numerical experiments on the FP SM identification
opment of simulation tools for technological processes (TP). on the example of a strict model of the deethanizer of an oil
Such tools are designed to create first principles (FP) models refinery’s typical gas fractionating unit (GFU).
of technological objects. These models allow reproducing in
detail the structure and characteristics of simulated units, tak- II. S IMULATION M ODEL I DENTIFICATION
ing into account the laws of conservation of mass and energy, FP SM of TP are nonlinear state space models. The follow-
as well as strict physical dependencies for thermodynamics, ing system of equations can generally describe their dynamics:
heat transfer, and fluid dynamics [1]. At the same time, in (
addition to the generally recognized areas of application of xi+1 = f i (xi , ui ; p) ,
high-precision simulation models of TP, such as the design (1)
y i = g i (xi , ui ) .
of technological units, determination of nominal equipment
characteristics, determination of “bottlenecks,” “what-if” anal-
   
(1) (k) (1) (l)
Here xi = xi , . . . , xi , y i = yi , . . . , yi , and
ysis, and operator training systems [2], there is an increase in  
(1) (m)
interest in using this model type for TP control tasks [3]. ui = ui , . . . , ui are values of the model state, output,
Traditionally, control systems based on predictive models and control vectors, correspondingly, at time ti from given
n
(MPC — model predictive control) use empirical models time grid T = {ti }i=0 ; p – model tuning parameter vector.
(EM), which represent dependencies (often linear) of output It can be seen from (1) that the quality of the forecast for a
signals from input signals obtained during observation of the given control depends both on the proximity of the initial state
modeling object. The advantages of such models are their of the model xi to the state of the actual technological object at
simplicity, high calculation speed, and a general theory that the corresponding time and the parameter vector p. Since any
allows the application of these models to different types of TP is always under the influence of a considerable number of
TP. In addition, in the case of linear models, the convergence various kinds of disturbances, its state is constantly changing.
of numerical methods of their calculation is guaranteed, which Therefore, some tools are needed to allow to bring a model to
means high reliability, which is an essential criterion for the the object’s current state in real-time. The vector of parameters
functioning of a control system. At the same time, the assump- can be considered a constant for short periods (for most oil
tions made when creating such models, and their structural refining processes, within about a month). However, during
inconsistency with the modeling object, limit the range of the operation cycle of a considered unit (the typical period of
process parameters significantly, in which EM provides high operation between scheduled repairs is about one year), these
accuracy in predicting the object’s behavior. In case of changes values also experience slow drift. Hence, they need periodic
in operating conditions or process parameters, the EM must adjustment.
be calibrated. That means that the object is affected by test Under the identification of the model, we will understand
signals (disturbances), which decrease the TP control quality. the simultaneous assessment of the current state and param-

979-8-3503-3790-7/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on February 28,2024 at 14:09:42 UTC from IEEE Xplore. Restrictions apply.
 eters of the considered model. For this, it is convenient to
introduce an extended state vector:
 
xi
ξi = (2)
p
Then, we can rewrite the system (1) in the following form:
(
ξ i+1 = φi (ξ i , ui ) ,
(3)
y i = ψ i (ξ i , ui ) .
In this form, the identification problem (i.e., simultaneous
estimation of the state and parameters) represents the classical
filtering problem, which can be solved, for example, using an
extended Kalman filter.
III. D ISTILLATION C OLUMN M ODEL
The main technological line of GFU consists of several
distillation columns (Fig. 1). Each column extracts a certain
fraction of the plant feed [4]. The number of columns in a
particular unit usually ranges from three to seven, depend-
ing on the processed raw materials and target products. All
columns have a similar configuration and differ in the number
of plates, geometric dimensions, and nominal values of the
normal process conditions.
On each tray, the vapor coming from the lower tray through
the holes passes through the layer of liquid flowing through the
downcomer from the higher tray and comes into contact with
it (Fig. 2). This leads to a mass exchange between vapor and
liquid on the tray, as a result of which the heaviest components
pass from vapor to liquid, and the lightest components from
liquid to vapor. Thus, the concentration of heavy fractions
gradually increases in the liquid moving down the plates from
top to bottom, while the content of light components increases
in the vapor rising from bottom to top.
Due to the finiteness of the residence time of vapor and
Fig. 1. Distillation column operating principle
liquid on the tray, the phases do not reach the absolute ther-
modynamic equilibrium. Therefore, an essential characteristic
of the functioning of the tray is its efficiency, which is a
measure of the approximation of an actual tray operation to a
theoretical tray, that is, a hypothetical tray, on which an ideal
equilibrium between the vapor and liquid phases occurs. At
an efficiency equal to 0%, the vapor and liquid composition
entering the tray do not change; at an efficiency value of 100%,
equilibrium compositions leave the tray. There are different
definitions for the tray efficiency: for example, as a general
(average) efficiency as a local efficiency (Murphree method);
there are also methods based on the mass transfer theory [5].
In this article, we will understand the efficiency of a tray
as the proportion of the tray substance that reaches an ideal
thermodynamic equilibrium. The efficiency of the tray depends
not only on the tray’s design features but also on the current
process parameters: component compositions, flow rates and
temperatures of streams entering the tray, and pressure on the
tray. This parameter in the column model must be updated Fig. 2. Distillation column tray construction
as the process conditions change. Also, the trays’ efficiency
decreases during the column’s operation due to contamination.

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on February 28,2024 at 14:09:42 UTC from IEEE Xplore. Restrictions apply.
 Thus, tray efficiencies are the critical parameters of a column
model and a column identification algorithm should update the
efficiencies.
We have chosen a deethanization column for numerical
identification experiments since it is the first column in the
GFU process chain. Therefore, changes in the characteristics
of raw materials entering the plant strongly influence the
column. Moreover, the deethanization column has the most
significant variability of the process parameters. The raw
material of the column is a mixture of stabilization heads of
oil refining units and liquefied petroleum gas. The products of
the column are ethane-propane mixture, which feeds the fuel
gas system of the plant, and natural gas liquids (NGL) with a
low ethane concentration.
IV. N UMERICAL E XPERIMENTS
There were some attempts to identify column models earlier,
e.g. in [6] and [7]. Nevertheless, all the works on this topic
contain significant assumptions in designing the model. This
article considers an example of the identification of a strict
column SM, in which the material and energy balances,
the geometry of the column, and the hydraulic equations
describing the flow of substance between trays are fully taken
into account.
We built two models for conducting experiments. The first
model was created in a specialized TP modeling software,
UniSim® Design (Fig. 3), and was used as a source of
synthetic input data (instead of a natural plant). Hereafter, we
will call this model as the “object.” We also implemented the
second model in Python and used it to study the identification
algorithm based on the extended Kalman filter.
The considered deethanizer consists of 43 trays. Feed fluid
flows to the 30th and the 35th trays (bottom-up numbering).
State vector x consists of k = 1672 components, including
the following physical characteristics of the reflux drum, the
reboiler, and each tray:
• Amount of each chemical component (Table I) in both
vapor and liquid;
• Molar enthalpy of both vapor and liquid;
• Temperature;
• Vapor compressibility factor;
• Liquid molar density. Fig. 3. Structure of GFU deethanizer model built in UniSim® Design
In addition, the state vector stores openings of the control
valves in the control loops below:
• Column top tray pressure;
• y (3) – reflux drum pressure;
• Temperature of the condenser outlet;
• y (4) – reflux drum liquid level;
• Reflux drum pressure;
• y (5) – reflux flow rate;
• Reflux drum liquid level;
• y (6) – column sump liquid level;
• Reflux flow rate;
th
• y (7) – bottom product flow rate;
• the 5 plate temperature;
• y (8) – column sump temperature;
• Column sump liquid level;
• y (9) – column sump pressure;
• Deethanized NGL flow rate.
• y (10) , . . . , y (15) – temperatures at the 43rd , 35th , 30th , 10th ,
Model outputs: 5th , and 1st trays, correspondingly;
(1)
• y – top tray pressure; • y (16) , y (17) – correspondingly fractions of propane in the
(2)
• y – temperature of the condenser outlet; fuel gas and ethane in the column bottom product.

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on February 28,2024 at 14:09:42 UTC from IEEE Xplore. Restrictions apply.
 TABLE I TABLE II
C HEMICAL C OMPONENTS U SED FOR S UBSTANCE D EFINITION IN THE S CALE FACTORS FOR M ODEL O UTPUTS
B UILT M ODELS
Quantity Units Scale
Component index Chemical compound Pressure kPa 20.0
0 Nitrogen Temperature K 0.5
1 Hydrogen Flow rate m3 /h 0.5
2 Methane Liquid level % 1.0
3 Ethane Fuel gas propane content fraction 0.01
4 Ethene NGL ethane content fraction 0.0002
5 Propane
6 Propene
7 Isobutane
8 Normal butane tuned model in the initial state and the tuned parameter
9 Trans-2-butene
10 2,2-dimethylpropane values to the model error);
11 Isopentane • Identification with noisy outputs of the object to deter-
12 Normal pentane mine the dependence of the identification quality on the
13 Trans-2-pentene
noise level;
14 Normal hexane
15 Normal heptane • Identification of a model with structural differences with
the object (we changed the geometry of the column in
the model).
The following types of disturbances are supported: For all launches, except for the variant without identifica-
tion, we determined the factor of reduction of the error relative
• Composition, temperature, and flow rate of feed supplied
to the launch without identification:
to the 30th and the 35th trays;
• Temperature and flow rate of feed supplied to the top tray ENoID
q= (5)
(the stream “EXTERNAL REFLUX” in Fig. 3); EID
• The flow rate of the gas for maintaining pressure in feed
Here ENoID is the error value for running without identifi-
tanks (the stream “GAS TO E509” in Fig. 3); cation, and EID is for running with identification.
• Pressures of boundary streams (“FUEL GAS,” The influence of identification on the approach of the tuned
“HC COND,” and “K1 TO K2” in Fig. 3). model output to the object is shown in Fig. 4 (the case without
We used tray efficiencies as the model identification tuning object noise) and Fig. 5 (the case of noisy object).
parameters with the assumption that the efficiencies of all trays
are equal.
V. R ESULTS
We carried out several series of numerical experiments. In
each series, we launched the “object” and the custom model
from some initial states (generally not matching) by applying
the same disturbances and control actions to both the “object”
and the model.
After each launch of the model, we calculated the dis-
crepancy between the model and the object according to the
following quadratic criterion:
n X l
!2
(j) (j)
1 X yi − zi
E= (4)
(n + 1)l i=0 j=1 sj
 
(1) (l)
Here z i = zi , . . . , zi – the object output vector,
sj (j = 1, . . . , l) – scale factors were chosen based on physical
interpretation and measurement units of a particular output
(Table II).
We run the tuned model several times under different Fig. 4. Changing NGL ethane content (object outputs without noise)
conditions (each time from the same state):
• Without identification to determine the base error for The experiments confirm the principle possibility of using
identification quality evaluation; the extended Kalman filter with moderate deviations of the
• Identification without object output noise (shows the tuned model’s initial state from the state of the object and noise
contribution of differences between the object and the amplitudes of the order of magnitudes indicated in Table II.

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on February 28,2024 at 14:09:42 UTC from IEEE Xplore. Restrictions apply.
 slower than in real-time, making it inapplicable for TP control
tasks.
The high dimension of the model causes both of these dis-
advantages. Currently, we could see some efforts on building
hybrid models of technological objects (e.g. [8] and [9]), which
would be close in accuracy to FP models while having smaller
dimensions and higher calculation speed. The model type
seems promising as a predictive model for a new generation
of MPC.
R EFERENCES
[1] D. Foo, Chemical Engineering Process Simulation, 2nd ed. Amsterdam:
Elsevier, 2022.
[2] J. Philip, Operator Training Simulator Handbook: Best practices for
developing and investing in OTS. Birmingham: Packt Publishing, 2022.
[3] R. Rhinehart, “Nonlinear model-predictive control using first-principles
models,” Smart Sensors Measurement and Instrumentation: Select Pro-
ceedings of CISCON 2021, S. Chokadi, R. Bandyopadhyay, Eds. Sin-
gapore: Springer, 2023, pp. 93–114.
[4] A. J. Kidnay, W. R. Parrish, and D. G. McCartney, Fundamentals of
Natural Gas Processing, 3rd ed. Boca Raton, FL: CRC Press, 2019.
[5] T. L. Domingues, A. R. Secchi, and T. F. Mendes, “Overall efficiency
evaluation of commercial distillation columns with valve and dualflow
Fig. 5. Changing NGL ethane content (noisy object) trays,” AIChE Journal, vol. 56, no. 9, pp. 2323–2330, Sept. 2010.
[6] D. R. Yang and K. S. Lee, “Monitoring of a distillation column using
modified extended Kalman filter and a reduced order model,” Computers
and Chemical Engineering, vol. 21, no. SUPPL.1, pp. S565-S570, 1997.
In addition, identification continued to be successful with [7] S. Dodds, G. Adams, W. Heath, and G. Goodwin, “Application of an
introducing errors (up to 10%) to the geometric parameters extended Kalman filter to a binary distillation column model,” IFAC
of the column in the tuned model. However, this reduced the Proceedings Volumes, vol. 34, no. 25, pp. 421–426, 2001.
[8] W. Bradley et al., “Perspectives on the integration between first-
quality of the forecast (Table III). Nevertheless, the difference principles and data-driven modeling,” Computers and Chemical Engi-
in the values of state variables or noise amplitudes becomes neering, vol. 166, 107898, Oct. 2022.
several times greater than the scales indicated in Table II. In [9] L. Rajulapati, C. Sivadurgaprasad, B. Shyamala, and R. Rengaswamy,
“Integration of machine learning and first principles model,” AIChE
that case, the model with the extended Kalman filter becomes Journal, vol. 56, no. 6, e17715, June 2022.
unstable. Furthermore, due to the high dimension of the
model, the Kalman filter significantly slows down calculation
since an estimate of partial derivatives overall state vector
components is required. The model’s acceleration running
without identification averages between 65 and 70 in real-
time, while with the Kalman filter, the acceleration becomes
0.12–0.15 (6 to 8 times slower than in real-time).

TABLE III
E RROR R EDUCTION FACTORS

Series no. No ID Noise Struct. diff.

1 8.10 4.22 5.48
2 7.32 3.94 4.81
3 7.67 4.15 5.39
4 7.91 4.26 5.29
5 7.48 4.05 5.07

VI. C ONCLUSION
The performed calculations confirm the fundamental possi-
bility of GFU distillation column FP SM identification. At the
same time, we have found limitations on the stability of the
utilized identification algorithm in cases of significant differ-
ences in the initial states of the tuned model and the object,
as well as with strong noise of the object outputs. In addition,
during the identification process, the model calculation became

Authorized licensed use limited to: Universidad Nacional de Colombia (UNAL). Downloaded on February 28,2024 at 14:09:42 UTC from IEEE Xplore. Restrictions apply.