See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/262575873

Design and Comparison of Constrained MPC With PID Controller for Heave
Disturbance Attenuation in Offshore Managed Pressure Drilling Systems

Article in Marine Technology Society Journal · March 2014

DOI: 10.4031/MTSJ.48.2.4

CITATIONS READS

19 432

4 authors:

Amirhossein Nikoofard Tor Arne Johansen

Norwegian University of Science and Technology Norwegian University of Science and Technology
56 PUBLICATIONS 254 CITATIONS 631 PUBLICATIONS 18,681 CITATIONS

SEE PROFILE SEE PROFILE

Hessam Mahdianfar Alexey Pavlov

University of Leicester Norwegian University of Science and Technology
15 PUBLICATIONS 206 CITATIONS 121 PUBLICATIONS 2,710 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Leak Detection View project

Production Optimization in Upstream Oil and Gas View project

All content following this page was uploaded by Hessam Mahdianfar on 12 January 2016.

The user has requested enhancement of the downloaded file.

Code:48.2002d

Design and Comparison of Constrained MPC with PID Controller for

Heave Disturbance Attenuation in Offshore Managed Pressure Drilling
Systems
Amirhossein Nikoofard, Tor Arne Johansen, Hessam Mahdianfar, and Alexey Pavlov

Abstract:
This paper presents a constrained finite horizon model predictive control (MPC) scheme for
regulation of the annular pressure in a well during managed pressure drilling from a floating
vessel subject to heave motion. In addition the robustness of a controller to deal with heave
disturbances despite uncertainties in the friction factor and bulk modulus is investigated. The
stochastic model describing sea waves in the North Sea is used to simulate the heave
disturbances. The results show that the closed-loop simulation without disturbance has a fast
regulation response, without any overshoot, and is better than a proportional-integral-
derivative (PID) controller. The constrained MPC for managed pressure drilling shows further
improved disturbance rejection capabilities with measured or predicted heave disturbance.
Monte Carlo simulations show that the constrained MPC has a good performance to regulate
set point and attenuate the effect of heave disturbance in case of significant uncertainties in
the well parameter values.

Keywords: Managed pressure drilling, heave compensation and model predictive control.

1. INTRODUCTION
In drilling operations, a drilling fluid (mud) is pumped down through the drill string
and flows through the drill bit at the bottom of the well (Figure 1). The mud flows up the well
annulus carrying cuttings out of the well. The mud is separated at the surface from the return
well flow, conditioned and stored in storage tanks (pits), before it is pumped down into the
well for further drilling. To avoid fracturing, collapse of the well, or influx of formation fluids
surrounding the well, it is crucial to control the pressure in the open part of the annulus within
a certain operating window. In conventional drilling, this is done by using a mud of
appropriate density and adjusting mud pump flow-rates. In managed pressure drilling (MPD),
the annulus is sealed and the mud exits through a controlled choke, allowing for faster and
more precise control of the annular pressure. In MPD operation, the dynamic pressure of the
well must be kept higher than the reservoir pore pressure to prevent gas or formation fluids
from entering the well, and less than a formation fracture pressure at all times 𝑡 and positions
𝑥:

𝑝𝑝𝑜𝑟𝑒 (𝑥) ≤ 𝑝𝑤𝑒𝑙𝑙 (𝑥, 𝑡) ≤ 𝑝𝑓𝑟𝑎𝑐 (𝑥) (1)

Code:48.2002d

where 𝑝𝑝𝑜𝑟𝑒 , 𝑝𝑤𝑒𝑙𝑙 , and 𝑝𝑓𝑟𝑎𝑐 are reservoir pore pressure, well pressure, and formation
fracture pressure, respectively. In automatic MPD systems, the choke is controlled to keep the
annular mud pressure between specified upper and lower limits. There are several studies
about different aspects of MPD modeling (e.g. see Landet et al. (2012a, 2013); Petersen et al.
(2008); Mahdianfar et al. (2013); Kaasa et al. (2012)). Estimation and control design in MPD
has been investigated by several researchers (e.g. see Kaasa et al.(2012); Nygaard et al.
(2007c); Breyholtz et al. (2010); Zhou et al. (2011); Zhou and Nygaard (2011); Godhavn et al.
(2011)). These studies are mainly focused on pressure control during drilling from a fixed
platform without any heave motion.

The automatic MPD system has several advantages compared to conventional drilling, as
follows:

 Reducing the drilling costs due to reducing the Non-Productive Time (NPT).
 Increasing the rate of penetration (ROP).
 Improving wellbore stability.
 Minimizing the risk of lost circulation.
 Extending control over Bottom-hole pressure (BHP) to operational scenarios such as
connections and trips and when the rig pumps are off.
 Improvement in safety and well control due to a more detailed design and planning
required for accomplishment.
Code:48.2002d

Figure 1. Schematic of an MPD system (Courtesy of Dr. Glenn-Ole Kaasa, Statoil

Research Centre.)
When designing MPD control systems, one should take into account various
operational procedures and disturbances that affect the pressure inside the well. There is a
specific disturbance occurring during drilling from floaters that significantly affects MPD
operations. In this case, the rig moves vertically with the waves, referred to as heave motion.
As drilling proceeds, the drill string needs to be extended with new sections. Thus, every
couple of hours or so, drilling is stopped to add a new segment of about 27 meters to the drill
string. During drilling, a heave compensation mechanism is active to isolate the drill string
from the heave motion of the rig. However, during connections, the pump is stopped and the
string is disconnected from the heave compensation mechanism and rigidly connected to the
rig. The drill string then moves vertically with the heave motion of the floating rig, and acts
like a piston on the mud in the well. The heave motion may be more than 3 meters in
amplitude and typically has a period of 10-20 seconds, which causes severe pressure
fluctuations at the bottom of the well. Pressure fluctuations have been observed to be an order
of magnitude higher than the standard limits for pressure regulation accuracy in MPD (about
±2.5 bar) (Godhavn(2010)). Downward movement of the drill string into the well increases
pressure (surging), and upward movement decreases pressure (swabbing). Excessive surge
Code:48.2002d

and swab pressures can lead to mud loss resulting from high pressure fracturing of the
formation or a kick-sequence (uncontrolled influx from the reservoir) that can potentially
grow into a blowout as a consequence of low pressure.

Rasmussen and Sangesland (2007) compared and evaluated different MPD methods
for compensation of surge and swab pressure. In Nygaard et al. (2007a), it is shown that surge
and swab pressure fluctuation in the bottom hole pressure during pipe connection can be
suppressed by controlling the choke and main pump. Nygaard et al. (2007b) used a nonlinear
model predictive control algorithm to obtain optimal choke pressure for controlling the
bottom-hole pressure during pipe connection in a gas dominant well. Pavlov et al. (2010)
presented two nonlinear control algorithms based on feedback linearization for handling
heave disturbances in MPD operations. Mahdianfar et al. (2012a, b) designed an infinite-
dimensional observer that estimates the heave disturbance. This estimation is used in a
controller to reject the effect of the disturbance on the down-hole pressure. In all the above
mentioned papers, the controllers are designed for the nominal case disregarding the
uncertainty in the parameters, though several parameters in the well could be uncertain during
drilling operations. In addition the heave disturbance, which is inherently stochastic and
contains many different harmonics, is approximated by one or a couple of sinusoidal waves
with known fixed frequencies throughout controller design and simulations. In this paper, a
stochastic model for the heave motion in the North Sea is given and is used in simulations.

Model predictive control (MPC) is one of the most popular controller design
methodologies for complex constrained multivariable control problems in industry and has
been the subject of many studies since the 1970s (e.g. see Mayne et al. (2000); Morari and
Lee (1999); Garcia et al. (1989); Maciejowski (2002)). At each sampling time, a MPC control
action is acquired by the on-line solution of a finite horizon open-loop optimal control
problem. Only the first part of the optimal control trajectory is applied to the system. At the
next sampling time, the computation is repeated with new measurements obtained from the
system. The purpose of this paper is to study a constrained MPC scheme for controlling the
pressure during MPD oil well drilling using measurements and optionally predictions of the
heave disturbances. In some cases short-term heave motion prediction based on forward-
looking sensors such as ocean wave radar may be (Kuchler et al. (2011a)) , and we can use
them directly in our MPC controller.. One of the criteria for evaluating the controller
performance is its ability to handle heave disturbances. This scheme is compared with a
standard proportional-integral-derivative (PID)-control scheme. Furthermore, the robustness
Code:48.2002d

of the controller to deal with heave disturbances despite significant uncertainties in the friction
factor and bulk modulus is investigated by Monte-Carlo simulations.

In the following sections, a model based on mass and momentum balances that
provides the governing equations for pressure and flow in the annulus is given. A stochastic
modeling of waves in the North Sea is used, and the heave disturbance induced by the
elevation motion of the sea surface is modeled. The design of a constrained MPC scheme is
presented and applied on MPD. In the cases with and without the predictive heave disturbance
feed-forward and prediction, it is shown that this controller outperforms a PID controller.
Finally robust performance of an MPC controller is evaluated through Monte-Carlo
simulations.

2. MATHEMATICAL MODELING

In this section, the distribution of single phase flows and pressures in the annulus and
the drill string is modeled by two coupled partial differential equations (PDE). Then, the PDE
model is discretized by using the finite volume method. Finally, the model describing the
vessel’s heave motion in response to the stochastic sea waves in the North Sea is presented
and used as the heave disturbance.

2.1 Annulus flow dynamics

The governing equations for flow in an annulus are derived from mass and momentum
balances based on one-dimensional hydraulic transmission line (Landet et al. (2013)).

𝜕𝑝(𝑥, 𝑡) 𝛽 𝜕𝑞(𝑥, 𝑡)
=− (2)
𝜕𝑡 𝐴(𝑥) 𝜕𝑥

𝜕𝑞(𝑥, 𝑡) 𝐴(𝑥) 𝜕𝑝(𝑥, 𝑡) 𝐹

=− − + 𝐴(𝑥)𝑔𝑐𝑜𝑠(𝛼(𝑥)) (3)
𝜕𝑡 𝜌0 𝜕𝑥 𝜌0

where 𝑝(𝑥, 𝑡) and 𝑞(𝑥, 𝑡) are the pressure and volumetric flow rate at location x and time t,
respectively. The bulk modulus of the mud is denoted by 𝛽. 𝐴(𝑥) is the cross section area, 𝜌
is the (constant) mass density, 𝐹 is the friction force per unit length, 𝑔 is the gravitational
constant and 𝛼(𝑥) is the angle between gravity and the positive flow direction at location 𝑥 in
the well (Figure 2). To derive a set of ordinary differential equations describing the dynamics
of the pressures and flows at different positions in the well, equations (2) and (3) are
discretized by using the finite volume method. To solve this problem, the annulus is divided
Code:48.2002d

into a number of control volumes, as shown in Figure 2, and integrating (2) and (3) over each
control volume. This model will be used for the MPC design.

Figure 2. Control volumes of annulus hydraulic model (Landent et al. (2013))

Landent et al. (2013) found that five control volumes could capture the main dynamics
of the system in the case of heave disturbance for a well from the Ullrigg test facility with a
particular length of about 2000 m and with water based mud (Landent et al. (2013)). Ullrigg is
a full scale drilling test facility located at the International Research Institute of Stavanger
(IRIS). The parameters corresponding to that well are used as a base case throughout this
paper. The set of nine ordinary differential equations describing five control volumes in the
annulus are as follows (Landet et al. (2012a,b))

𝛽1
𝑝1̇ = (−𝑞1 − 𝑣𝑑 𝐴𝑑 ) (4)
𝐴1 𝑙1

𝛽2
𝑝2̇ = (𝑞 − 𝑞2 ) (5)
𝐴2 𝑙2 1

𝛽3
𝑝3̇ = (𝑞 − 𝑞3 ) (6)
𝐴3 𝑙3 2

𝛽4
𝑝4̇ = (𝑞 − 𝑞4 ) (7)
𝐴4 𝑙4 3
Code:48.2002d

𝛽5
𝑝5̇ = (𝑞 − 𝑞𝑐 + 𝑞𝑏𝑝𝑝 ) (8)
𝐴5 𝑙5 4

𝐴𝑖 𝐹𝑖 (𝑞𝑖 )𝐴𝑖 Δℎ𝑖

𝑞̇ 𝑖 = (𝑝 − 𝑝𝑖+1 ) − − 𝐴𝑖 𝑔 (9)
𝜌𝑖 𝑙𝑖 𝑖 𝑙𝑖 𝜌𝑖 𝑙𝑖

𝑞𝑐 = 𝐾𝑐 √𝑝𝑐 − 𝑝0 𝐺(𝑢) (10)

where, 𝑖 = 1, . . . , 4, and the numbers 1, ..., 5 refer to the control volume number, with 1
being the lower-most control volume representing the down hole pressure (𝑝1 = 𝑝𝑏𝑖𝑡 ), and 5
being the upper-most volume representing the choke pressure (𝑝5 = 𝑝𝑐 ). 𝑣𝑑 is the heave
(vertical) velocity due to ocean waves and 𝐴𝑑 is the drill string cross section area. The length
of each control volume is denoted by 𝑙, and the height difference is ∆ℎ𝑖 . Since the well may
be non-vertical, 𝑙𝑖 and ∆ℎ𝑖 in general can differ from each other. The means for pressure
control are the backpressure pump flow 𝑞𝑏𝑝𝑝 and the choke flow𝑞𝑐 . The flow from the back
pressure pump 𝑞𝑏𝑝𝑝 is linearly related to the pump frequency and cannot be changed fast
enough to compensate for the heave-induced pressure fluctuations. Therefore, it is the choke
flow that is used primarily for control, and that is modeled by nonlinear orifice equation (10).
𝐾𝑐 is the choke constant corresponding to the area of the choke and the density of the drilling
fluid. 𝑝0 is the (atmospheric) pressure downstream of the choke and 𝐺(𝑢) is a strictly
increasing and invertible function relating the control signal to the actual choke opening,
taking its values on the interval [0, 1].

Based on experimental results from full scale tests at Ullrigg, the friction force in the
annulus is considered to be a linear function of the flow rate (Landet et al. (2013)). Friction
force on the 𝑖𝑡ℎ control volume is approximately modeled as

𝑘𝑓𝑟𝑖𝑐 𝑞𝑖
𝐹𝑖 (𝑞𝑖 ) = (11)
𝐴𝑖

where 𝑘𝑓𝑟𝑖𝑐 is constant friction coefficient.

Some components of the transient hydraulic model, (1)-(2), have significant uncertainties,
such as

 Rheology and viscosity of drilling fluid. Most drilling fluids are non-Newtonian, i.e.
with a nonlinear relation between shear stress and shear rate. Consequently, the
viscosity will not be constant over a cross-sectional flow area. To measure the shear
Code:48.2002d

stress/shear rate relationship, the viscometer measurements must be correlated with the
rheological model applied. However, information is limited and normally inadequate
for a model of high accuracy, particularly for modern oil based muds. Also, viscosity
may depend on pressure and temperature. Manual rheology measurements are
normally performed periodically on the rig at the atmospheric pressure and
temperature of the mud in the pit. Thus, information on the influence of temperature
and pressure variations is missing, (Lohne et al. (2008); Florence et al. (2010); Commented [AN1]: Missed information lead to inaccuracy
and uncertainty in the final model
Gravdal et al. (2010)). Commented [AJ2]: How important is this?

 Frictional pressure loss models for drill-pipe and annulus. The frictional pressure loss
depends on the mean cross sectional velocity, drilling fluid viscosity, flow regime, the
hydraulic diameter, and pipe roughness. The accuracy of all these derived parameters
is questionable. Moreover, the Fanning friction factor is a function of Reynolds
number where the Reynolds number is a function of the fluid viscosity for a
characteristic diameter (Kaasa et al. (2012); Florence et al. (2010); Lohne et al.
(2008)).
 Effective bulk modulus. A bulk modulus is used because the degree of mechanical
compliance of casing, pipe, hoses, and other components is uncertain and also it is
impossible to predict the amount of gas pockets, bubbles, or breathing of the well
(Kaasa et al. (2012)).

2.2 Waves Response Modeling

Environmental forces in the vertical direction due to waves are considered

disturbances to the motion control system of floating vessels. These forces, which can be
described in stochastic terms, are conceptually separated into low-frequency (LF) and wave- Commented [AJ3]: Why is the LF component not discussed
further in this section 2.2? Seems to me that some statement
frequency (WF) components (Fossen (2011)).The LF part is not considered any farther since should be made as to why this is not considered -- for
completeness.
it is very slow compared to the dynamics of the mud circulation system and dealt with by
other controllers and operationally (e.g. pipe connection ) .

During normal drilling operations the WF part of the drill-string motion is

compensated by the heave control system (Korde (1998); Do and Pan (2008); Kuchler et al.
(2011b)). However, during connections the drill-string is disconnected from the heave
compensation mechanism and rigidly connected to the rig. Thus, it moves vertically with the
heave motion of the floating rig and causes severe down-hole pressure fluctuations.

2.2.1 Linear Approximation for WF Position

Code:48.2002d

When simulating and testing feedback control systems, it is useful to have a simple
and effective way of representing the wave forces. Here the motion Response Amplitude
Operators (RAOs) are represented as a state-space model where the wave spectrum is
approximated by a linear filter. In this setting the RAO vessel model is represented in Figure
3, where 𝐻𝑟𝑎𝑜 (𝑠) is the wave amplitude-to-force transfer function and 𝐻𝑣 (𝑠) is the force-to-
motion transfer function. In addition to this, the response of the motion RAOs and the linear
vessel dynamics in cascade is modeled as constant tunable gains (Fossen (2011)). This means
that the RAO vessel model is approximated as (Figure 3)

𝐾 = 𝑑𝑖𝑎𝑔{𝐾1 , 𝐾 2 , 𝐾 3 , 𝐾 4 , 𝐾 5 , 𝐾 6 } (12)

𝐻𝑟𝑎𝑜 (𝑠) 𝐻𝑣 (𝑠) ≈ 𝐾 (13)

Since the vessel is typically designed to avoid resonances in the dominant wave
frequency, the fixed-gain approximation (equation (13)) produces good results in a closed-
loop system where the purpose is to test robustness and performance of a feedback control
system in the presence of waves.

Figure 3. Transfer function approximation for computation of wave-induced positions.

Then, the generalized WF position vector 𝜂𝑤 in Figure 3 becomes

𝜂𝑤 = 𝐾 𝐻𝑠 (𝑠) 𝑤(𝑠) (14)

where 𝐻𝑠 (𝑠) is a diagonal matrix containing transfer function with the spectral factors of the
wave spectrum 𝑆(𝜔). The WF position for the degree of freedom related to the heave motion
becomes

ℎ
𝜂𝜔 = 𝐾ℎ 𝜉ℎ (15)

𝜉 ℎ (𝑠) = ℎℎ (𝑠) 𝑤 ℎ (𝑠) (16)

where ℎℎ (𝑠) is the spectral factor of the wave spectral density function 𝑆(𝜔) and 𝑤 ℎ (𝑠) is a
zero-mean Gaussian white noise process with unity power across the spectrum:
Code:48.2002d

ℎ
𝑃𝑤𝑤 (𝜔) = 1.0 (17)

Hence, the power spectral density (PSD) function for 𝜉 ℎ (𝑠) can be computed as

ℎ
𝑃𝜉𝜉 (𝜔) = |ℎℎ (𝑗𝜔)|2 𝑃𝑤𝑤
ℎ (𝜔)
= |ℎℎ (𝑗𝜔)|2 (18)

2.2.2 JONSWAP Spectrum

The JONSWAP formulation is based on an extensive wave measurement program

known as the Joint North Sea Wave Project carried out in 1968 and 1969 in the North Sea,
between the island Sylt in Germany and Iceland. The JONSWAP spectrum is representative
of wind-generated waves under the assumption of finite water depth and limited fetch (Fossen
(2011); Ochi (2005)). The spectral density function is written

𝐻𝑠2 −5
−944 −4
𝑆(𝜔) = 155 4 ( 𝜔 exp( 𝜔 )) 𝛾 𝑌 (19)
𝑇1 𝑇14

where 𝐻𝑠 is the significant wave height, 𝑇1 is the average wave period, 𝛾 = 3.3 and

0.191𝜔𝑇1 − 1
𝑌 = exp[−( )2 ] (20)
√2𝜎

where

0.07 𝑓𝑜𝑟 𝜔 ≤ 5.24/𝑇1

𝜎={ (21)
0.09 𝑓𝑜𝑟 𝜔 > 5.24/𝑇1

The modal period, 𝑇0 , is related to the average wave period through 𝑇1 = 0.834 𝑇0
(Fossen (2011)).

Figure 4, which is produced using the Marine Systems Simulator (MSS) Toolbox,
shows the JONSWAP spectrum power distribution curve. The parameter values for 𝐻𝑠 and 𝑇0
are taken from Michel (1999). From Figure 4 we can see that the JONSWAP spectrum is a
narrow band spectrum, and its energy is mainly focused on 0.5 - 1.5 rad/s, and the peak
frequency is 𝜔0 = 0.7222 rad/s.
Code:48.2002d

Figure 4. JONSWAP spectrum and its approximation.

2.2.3 Second-Order Wave Transfer Function Approximation

As discussed earlier, a finite dimensional rational transfer function wave response

approximation for 𝐻𝑠 (𝑠) is usually preferred by ship control systems engineers, because of its
simplicity and applicability:

2𝜆𝜔0 𝜎𝑠
ℎℎ (𝑠) = (22)
𝑠 2 + 2𝜆𝜔0 𝑠 + 𝜔02

where 𝜆 = 0.1017, 𝜎 = 1.9528, 𝐻𝑠 = 4.70, 𝑇0 = 8.70, 𝜔0 = 0.7222 and 𝐾 ℎ = 9.1 are

typical parameters for heave motion of the drilling rig. The transfer function approximation is
shown in Figure 4.

3. CONTROLLER DESIGN

The model described by equations (4)-(10) is in the form of a nonlinear strict feedback
𝛽𝑗 𝐴𝑗
system, with an unmatched stochastic disturbance. By considering 𝑎𝑗 = , 𝑏𝑗 = , and
𝐴𝑗 𝑙𝑗 𝜌 𝑗 𝑙𝑗

𝐾𝑓𝑟𝑖𝑐
𝑐𝑗 = , the model in state-space form would be
𝜌𝑗 𝑙𝑗

𝑋̇ = 𝐴𝑋 + 𝐵 𝑢𝑎 + 𝐵1 + 𝐸 𝑑
{ (23)
𝑦 = 𝐶 𝑋,
Code:48.2002d

where

𝑋 = [𝑝1 𝑞1 𝑝2 𝑞2 𝑝3 𝑞3 𝑝4 𝑞4 𝑝5 ]𝑇

é 0 -a 0 0 0 0 0 0 0 ù
ê 1
ú
ê b1 -c1 -b1 0 0 0 0 0 0 ú
ê ú
ê 0 a2 0 -a2 0 0 0 0 0 ú
ê 0 0 b -c2 -b2 0 0 0 0 úú
ê 2

A= ê 0 0 0 a3 0 -a3 0 0 0 ú
ê ú
ê 0 0 0 0 b3 -c3 -b3 0 0 ú
ê ú
ê 0 0 0 0 0 a4 0 -a4 0 ú
ê 0 0 0 0 0 0 b4 -c4 -b4 úú
ê
ê 0 0 0 0 0 0 0 a5 0 ú
ë û
B  0 0 0 0 0 0 0 0 a5 
T
(24)
B1  263.7814  0 1 0 1 0 1 0 1 0
T

E   22.0857 0 0 0 0 0 0 0 0
T

C  1 0 0 0 0 0 0 0 0 
and

𝑢𝑎 = 𝑞𝑏𝑝𝑝 − 𝑞𝑐 (25)

The output 𝑦 = 𝑝1 is the bottom-hole pressure. The heave disturbance 𝑣𝑑 in equation

(4) will be compensated by using constrained MPC as designed in section 3.1. Note that the
hydrostatic pressures in equation (9) are included in the states 𝑝𝑖 in (4)-(8).

3.1 MPC

The main MPC objective in this paper is to regulate bottom-hole pressure to desired values
(set points) during pipe connection by minimizing the cost function and satisfying output and
input constraints.

3.1.1 Constrained MPC design

Consider the discrete-time linear time-invariant input-affine system (23) while

fulfilling the constraints

𝑦𝑚𝑖𝑛 ≤ 𝑦(𝑘) ≤ 𝑦𝑚𝑎𝑥 , 𝑢𝑚𝑖𝑛 ≤ 𝑢𝑎 (𝑘) ≤ 𝑢𝑚𝑎𝑥 (26)

Code:48.2002d

at all-time instants 𝑘 ≥ 0.

In (23)-(26), 𝑛, 𝑝 and 𝑚 are the number of states, outputs and inputs, respectively, and
𝑋(𝑘) ∈ ℜ𝑛 , 𝑦(𝑘) ∈ ℜ𝑝 , 𝑑(𝑘) ∈ ℜ𝑛 and 𝑢(𝑘) ∈ ℜ𝑚 are the state, output, disturbance and
input vectors, respectively.

The constrained MPC solves a constrained optimal regulation problem at each time k.


min {J (u, y, r )   [(u (k  i | k )T R u (k  i | k )  u (k  i | k )T R u u (k  i | k )
U {uk ,......,uk  N } i 1

 ( y (k  i | k )  r (k  i | k ))T Q ( y (k  i | k )  r (k  i | k )]}

𝑠𝑢𝑏𝑗𝑒𝑐𝑡 𝑡𝑜 𝑦𝑚𝑖𝑛 ≤ 𝑦𝑖+𝑘|𝑘 ≤ 𝑦𝑚𝑎𝑥 𝑖 = 1, … , 𝑁,

𝑢𝑚𝑖𝑛 ≤ 𝑢𝑖+𝑘|𝑘 ≤ 𝑢𝑚𝑎𝑥 𝑖 = 1, … , 𝑁, (27)

∆ 𝑢𝑚𝑖𝑛 ≤ ∆𝑢𝑖+𝑘|𝑘 ≤ ∆𝑢𝑚𝑎𝑥 𝑖 = 1, … , 𝑁,

𝑥𝑘|𝑘 = 𝑥(𝑘)

𝑥𝑖+𝑘+1|𝑘 = 𝐴𝑑𝑖 𝑥𝑖+𝑘|𝑘 + 𝐵𝑑𝑖 𝑢𝑖+𝑘|𝑘 + 𝐵𝑑𝑖,1 + 𝐸𝑑𝑖 𝑑𝑖+𝑘|𝑘

𝑦𝑖+𝑘|𝑘 = 𝐶𝑑𝑖 𝑥𝑖+𝑘|𝑘

where 𝑁, 𝐽 and 𝑟 are the finite horizon, cost function and reference trajectory, respectively.
The matrices 𝐴𝑑𝑖 , 𝐵𝑑𝑖 , 𝐵𝑑𝑖,1 , 𝐸𝑑𝑖 and 𝐶𝑑𝑖 follow from a discretization of the system. The
subscript "(𝑘 + 𝑖|𝑘)" denotes the value predicted for time 𝑘 + 𝑖, and it is assumed that 𝑄,
𝑅𝛿𝑢 and 𝑅 are the positive definite matrices.

Since the states 𝑥(𝑘) are not directly measurable, predictions are computed from
estimation of states. Since the pair (𝐶, 𝐴) is detectable, a state observer is designed to provide
estimation of states 𝑥(𝑘) as described in section 3.2. The controller computes the optimal
solution 𝑈 by solving the quadratic programing (QP) problem (27). If the future value of
disturbances and/or measurement of disturbances are not assumed to be known then
disturbances are assumed to be zero in the MPC predictions.

Controller parameters such as weight of inputs, inputs rate and outputs and control
horizon must be tuned to achieve the good performance and stability in this problem. The
Code:48.2002d

prediction horizon should be chosen large enough to ensure the closed-loop stability of the
control system.

3.1.2 MPC Constraints

The upper and lower bounds on the input are chosen from the choke opening modes,
which are fully opened and fully closed, respectively. Enforcing pressure of the annulus in a
certain operating window is the main reason for using MPD. The hydrostatic pressure of the
well must be kept between both the reservoir formation pressure and collapsing pressure on
one side and fracturing pressure on the other side. The typical limits for pressure regulation
accuracy in MPD is about ±2.5 bar. The controlled output constraints for the limits for
pressure regulation accuracy in MPD must be softened by the addition of slack variables.

3.1.3 MPC Cost Function

The cost function (28) consists of three standard terms. The first term penalizes the
prediction input effort and the second term in the cost function penalizes variation in the
prediction control input. The last term weights the deviations of the output variable from the
reference trajectory 𝑟(𝑘 + 𝑖|𝑘).

3.2. Kalman Filter for state estimation

The discrete-time Kalman filter is a recursive algorithm based on discrete linear

dynamic systems and known stochastic models of noise and disturbance. The Kalman filter
has ability to estimate states with the minimum variance of the estimation error. This
algorithm has two distinct steps: prediction and correction. In the prediction step, predicted
state ( 𝑥̂𝑘|𝑘−1 ) and predicted estimate covariance (𝑃𝑘|𝑘−1 ) are computed. In the correction step
with updated measurement, optimal Kalman gain (𝐾𝑘 ) is computed. Then, updated state (𝑥̂𝑘|𝑘 )
and updated estimate covariance (𝑃𝑘|𝑘 ) are computed with optimal Kalman gain. More details
on Kalman filtering can be found in Simon (2006).

4. SIMULATION RESULTS

The nominal parameters for simulations, identified from the IRIS Drill simulator (Nygaard et
al. 2007d), are given in Table 1.

Table 1.Parameter Values

Code:48.2002d

Parameter Value Parameter Value

a 2.254× 108 [Pa/m3] g(gravitational constant) 9.806[m/s2]
b 4.276×10-8 [m4/Kg] A (annulus area) 0.0269 [m2]
Kf (friction coefficient) 5.725×105 [sPa/m3] Ad (drill string area) 0.0291 [m2]
qbpp ( backpressure pump 369.2464 [m3/s] Kc (choke constant) 2.32
flow)
c 14.4982 [1/sm2] p0 (atmospheric pressure) 101325[pa]
The time-step used for discretizing the dynamic optimization model was 0.1 s. This is
also the sampling interval of measurements and the update prediction of the Kalman filter and
MPC. The input weight (𝑅), input rate weight (𝑅𝛿𝑢 ), output weight (𝑄) and prediction horizon
(𝑁) are chosen as 150, 0, 17 and 100, respectively. In this problem, the prediction horizon
(10s) is relatively large compared with the settling time to ensure the closed-loop stability of
the control system. The weights specify trade-offs in the controller design. Choosing a larger
output weight or smaller input weight results in overshoot in the closed-loop response and
sometimes broken constraints. On the other hand, if a larger input weight or smaller output
weight are chosen, then the closed loop response is slower or sometimes unstable.

To compare the impact of MPC on the drilling system with other controllers, a PID
controller was applied to the system as well. A PID controller is chosen due to its popularity
in the industry. Proportional, integral and derivative gains are chosen as 0.75, 0.002 and -1,
respectively. The Bode plot of the loop transfer function with the PID is shown in Figure 5.
Bandwidth with PID is less than 1.3 rad/sec, and the phase drops very quickly. Therefore, it is
not realistic to get a bandwidth of about 5 rad/sec or more, as would be desirable for this
disturbance which has dominating frequencies of about 0.5 - 1.5 rad/s.
Code:48.2002d

Figure 5. Bode plot of the loop transfer function with the PID.

Several simulations are performed. The first simulation is shown in Figure 6, where
the nominal model is used for generating the measurements, and there is no heave motion. A
soft constraint of ±2.5bar (compared to the reference pressure) and a constraint of choke
opening taking its values on the interval [0,1] are included in the constrained MPC
optimization. Figure 6 compares the responses of the PID controller and constrained MPC to
regulate a set point trajectory. In the proposed MPC controller, the bottom-hole pressure
approaches to set point quickly without any overshoot. In comparison to the MPC controller,
the PID controller has some overshoot and a slower response. The choke control signal in the
constrained MPC is illustrated in Figure 6 (b).
Code:48.2002d

Figure 6.(a) Bottom-hole pressure without disturbance.

(b) MPC control signal to the choke without disturbance

The second simulation is shown in Figures 7 and 8, where the nominal model with
heave disturbance is used for generating the measurements. The same constraints as in the
previous simulation are enforced to the controller. Figure 7 compares the responses of
constant input (𝑞𝑏𝑝𝑝 = 𝑞𝑐 ) and constrained MPC to track the set point reference with existing
heave disturbance. A constant input could not reduce the effect of heave disturbance and track
the set point reference. Figure 8 (a) compares the responses of PID controller and constrained
MPC to track the set point reference with a heave disturbance. It is found that the MPC
controller is capable of maintaining the constraints whereas the PID controller is not.
Performance of the controller is evaluated through the root mean square (RMS) tracking error
metric. The RMS tracking errors for the MPC and the PID controller are 1.2524 and 1.6273,
respectively, which means that the effect of disturbances is reduced to 77.0% by the MPC
Code:48.2002d

compared to the PID. As indicated in this figure and RMS tracking error, the constrained
MPC shows good disturbance rejection capabilities. The choke control signal is illustrated in
Figure 8 (b). Figure 9 shows the heave disturbance pressure variations.

Figure 7. Bottom-hole pressure

Code:48.2002d

Figure 8.(a) Bottom-hole pressure with heave disturbance.

(b) MPC control signal to the choke with heave disturbance

Figure 9. Heave disturbance.

The next simulation is shown in Figure 10 where the nominal model with heave
disturbance is used for generating the measurements. The same constraints as in the previous
simulation are enforced to the controller. In this simulation, the heave disturbance is assumed
to be predictable. The heave disturbance is given by 𝑣𝑑 = 𝑐𝑜𝑠(2𝜋𝑡/12)[𝑚], where 2𝜋/12
corresponds closely to the most dominant wave frequency in the North Atlantic, with
reference to the JONSWAP spectrum (Landet et al. (2012b, 2013)). The input weight for
MPC with future knowledge of heave disturbance is chosen as 𝑅 = 85. Figure 10 compares
the responses of the MPC controller without future knowledge of heave disturbance and the
MPC with future knowledge of heave disturbance to track the set point reference. It is found
that the MPC controller with future knowledge of heave disturbance reduces the effect of
Code:48.2002d

heave disturbance more significantly than MPC controller without future knowledge of heave
disturbance. The MPC can therefore efficiently utilize predictions of future heave disturbance
to improve the control.

Figure 10. Bottom-hole pressure with predictable heave disturbance.

4.1. Robustness analysis of closed-loop system using Monte-Carlo simulations

Finally, the robustness of the constrained MPC without future knowledge of heave
disturbance with the presence of uncertainties in the friction factor and bulk modulus, 25%
each, is evaluated by Monte-Carlo simulations. Each simulation time was 200 seconds and the
simulations were done over 400 Monte-Carlo runs in the uncertainty region with uniform
distribution. We evaluated the performance by computing the ratio of average of RMS of the
tracking error to RMS of the stochastic disturbance. The result indicates that the controller
was successful to attenuate the disturbance in the uncertain system to 74.34%.

5. CONCLUSIONS

In this paper a dynamical model describing the flow and pressure in the annulus is
used. The model was based on a hydraulic transmission line, and is discretized using a finite
volume method. The disturbance due to drill-string movement is simulated as a stochastic
model describing sea waves in the North Sea applied to the flow in the bottom-hole of the
well.
Code:48.2002d

A constrained MPC for controlling bottom-hole pressure during oil well drilling was
designed. It was found that the constrained MPC scheme is able to successfully control the
down-hole pressure. It was also found that a constrained MPC shows improved attenuation of
the heave disturbance. Comparing the PID controller results with MPC shows that the MPC
controller has a better performance than the PID controller, being able to reduce the effect of
disturbances to 77%. Monte Carlo simulations show that the constrained MPC has a good
performance to regulate the set point and attenuate the effect of the heave disturbance in case
of significant uncertainties in the well parameter values. Finally, it is shown that performance
can be further improved by prediction of the heave motion about 10 seconds into the future.

ACKNOWLEDGMENT

The authors gratefully acknowledge the financial support provided to this project through the
Norwegian Research Council and Statoil ASA (NFR project 210432/E30 Intelligent Drilling).

REFERENCES:

Breyholtz, Ø., Nygaard, G., Siahaan, H., Nikolaou, M., 2010. Managed pressure drilling: A
multi-level control approach. In: SPE Intelligent Energy Conference and Exhibition. No.
128151-MS. Society of Petroleum Engineers, Utrecht, The Netherlands.

Do, K., Pan, J., 2008. Nonlinear control of an active heave compensation system. Ocean
Engineering 35 (5-6):558-571.

Florence, F., Iversen, F., 2010. Real-Time Models for Drilling Process Automation: Equations
and Applications. In IADC/SPE Drilling Conference and Exhibition

Fossen, T. I., 2011. Handbook of Marine Craft Hydrodynamics and Motion Control. John
Wiley &Sons. 600 pp.

Garcia, C. E., Prett, D. M., Morari, M., 1989. Model predictive control: Theory and practice-a
survey. AUTOMATICA 25 (3):335-348

Godhavn, J. M., (2010). Control Requirements for Automatic Managed Pressure Drilling
System. SPE Drilling and Completion 25(3): 336-345.
Code:48.2002d

Godhavn, J.-M., Pavlov, A., Kaasa, G.-O., Rolland, N. L., 2011. Drilling seeking automatic
control solutions. In Proceedings of the 18th World Congress. Vol. 18. The International
Federation of Automatic Control, IFAC, Milano, Italy, pp. 10842-10850.

Gravdal, J., Lorentzen, R., Fjelde, K. K., Vefring, E., 2010. Tuning of computer model
parameters in managed-pressure drilling applications using an unscented-kalman-filter
technique. SPE Journal, 15(3):856-866.

Kaasa, G.-O., Stamnes, Ø. N., Aamo, O. M., Imsland, L. S., 2012. Simplified hydraulics
model used for intelligent estimation of down-hole pressure for a managed pressure-drilling
control system. SPE Drilling and Completion 27 (1):127-138.

Korde, U. A., 1998. Active heave compensation on drill ships in irregular waves. Ocean
Engineering 25 (7):541-561.

Kuchler, S., Eberharter, J. K., Langer, K., Schneider, K., & Sawodny, O. ,2011a. Heave
Motion Estimation of a Vessel Using Acceleration Measurements. Proceedings of the 18th
World Congress. Vol. 18. The International Federation of Automatic Control, IFAC, Milano,
Italy, pp. 14742-14747.

Kuchler, S., Mahl, T., Neupert, J., Schneider, K., Sawodny, O., 2011b. Active control for an
offshore crane using prediction of the vessel's motion. IEEE/ASME Transactions on
Mechatronics 16 (2):297-309.

Landet, I. S., Mahdianfar, H., Aarsnes, U. J. F., Pavlov, A., Aamo, O. M., 2012a. Modeling
for mpd operations with experimental validation. In: IADC/SPE Drilling Conference and
Exhibition. No. SPE-150461. Society of Petroleum Engineers, San Diego, California.

Landet, I. S., Pavlov, A., Aamo, O. M., 2013. Modeling and control of heave-induced
pressure fluctuations in managed pressure drilling. IEEE Transactions on Control Systems
Technology.

Landet, I. S., Pavlov, A., Aamo, O. M., Mahdianfar, H., 2012b. Control of heave-induced
pressure fluctuations in managed pressure drilling. In: American Control Conference (ACC).
IEEE, Montreal, Canada, pp. 2270-2275.
Code:48.2002d

Lohne, H. P., Gravdal, J. E., Dvergsnes, E., Nygaard, G., Vefring, E., 2008 . Automatic
Calibration of Real-Time Computer Models in Intelligent Drilling Control Systems-Results
From a North Sea Field Trial. In International Petroleum Technology Conference.

Maciejowski, J. M., 2002. Predictive Control with Constraints. Prentice Hall. 352 pp.

Mahdianfar, H., Aamo, O. M., Pavlov, A., 2012a. Attenuation of heave-induced pressure
oscillations in offshore drilling systems. In: American Control Conference (ACC).IEEE,
Montreal, Canada, pp. 4915-4920.

Mahdianfar, H., Aamo, O. M., Pavlov, A., 2012b. Suppression of heave-induced pressure
fluctuations in MPD. In: Proceedings of the 2012 IFAC Workshop on Automatic Control in
Offshore Oil and Gas Production. Vol. 1. IFAC, Trondheim, Norway, pp. 239-244.

Mahdianfar, H., Pavlov, A., Aamo, O. M., 2013. Joint unscented kalman filter for state and
parameter estimation in managed pressure drilling. In: European Control Conference. IEEE,
Zurich, Switzerland.

Mayne, D. Q., Rawlings, J. B., Rao, C. V.and Scokaert, P. O., 2000. Constrained model
predictive control: Stability and optimality. AUTOMATICA 36 (6):789-814.

Michel, W. H., 1999. Sea spectra revisited. Marine Technology 36 (4):211-227.

Morari, M., Lee, J. H., May 1999. Model predictive control: past, present and future.
AUTOMATICA 23 (4-5):667-682.

Nygaard, G., Johannessen, E., Gravdal, J. E., Iversen, F., 2007a. Automatic coordinated
control of pump rates and choke valve for compensating pressure fluctuations during surge
and swab operations. Number 108344-MS, IADC/SPE Managed Pressure Drilling and
Underbalanced Operations Conference and Exhibition.

Nygaard, G., Vefring, E., Fjelde, K., Nævdal, G., Lorentzen, R., Mylvaganam, S, 2007b.
Bottomhole pressure control during drilling operations in gas-dominant wells. SPE Journal,
12(1):49-61.

Nygaard, G., Imsland, L, Johannessen, E., 2007c. Using nmpc based on a low-order model for
controlling pressure during oil well drilling. In 8th International IFAC Symposium on
Dynamics and Control of Process Systems, Volume 1, pp. 159-164. Mexico.
 Code:48.2002d

Nygaard, G., Gravdal, J., 2007d. Wemod for matlab user’s guide. International Research
Institute of Stavanger AS (IRIS), Bergen, Norway, Tech.Rep. 2007/234.

Ochi, M. K., 2005. Ocean Waves - The Stochastic Approach. Cambridge University Press.
332 pp.

Pavlov, A., Kaasa, G.-O., Imsland, L., 2010. Experimental disturbance rejection on a full-
scale drilling rig. In: 8th IFAC Symposium on Nonlinear Control Systems.IFAC, Bologna,
Italy, pp. 1338-1343.

Petersen, J., Rommetveit, R., Bjørkevoll, K. S., Frøyen, J., 2008. A general dynamic model
for single and multi-phase flow operations during drilling, completion, well control and
intervention. In: IADC/SPE Asia Pacific Drilling Technology Conference and Exhibition.
No.114688-MS. Society of Petroleum Engineers, Jakarta, Indonesia.

Rasmussen, O. S., Sangesland, S., 2007. Evaluation of mpd methods for compensation of
surge-and-swab pressures in floating drilling operations. No. 108346-MS.IADC/SPE
Managed Pressure Drilling & Underbalanced Operations, IADC/SPE, Texas, U.S.A.

Simon, D., 2006. Optimal state estimation: Kalman, H infinity, and nonlinear approaches.
Wiley. com, 530 pp.

Zhou, J., Nygaard, G., 2011. Automatic model based control scheme for stabilizing pressure
during dual gradient drilling. Journal of Process Control 21 (8):1138-1147.

Zhou, J., Stamnes, Ø. N., Aamo, O. M., Kaasa, G.-O.,2011. Switched control for pressure
regulation and kick attenuation in a managed pressure drilling system. IEEE Transactions on
Control Systems Technology 19 (2):337-350.

View publication stats