Clemson University

TigerPrints
All Dissertations Dissertations

8-2018

Model Predictive Control of Modern High-

Degree-of-Freedom Turbocharged Spark Ignited
Engines with External Cooled EGR
Rohit Vishvanath Koli
Clemson University, [email protected]

Follow this and additional works at: https://tigerprints.clemson.edu/all_dissertations

Recommended Citation
Koli, Rohit Vishvanath, "Model Predictive Control of Modern High-Degree-of-Freedom Turbocharged Spark Ignited Engines with
External Cooled EGR" (2018). All Dissertations. 2211.
https://tigerprints.clemson.edu/all_dissertations/2211

This Dissertation is brought to you for free and open access by the Dissertations at TigerPrints. It has been accepted for inclusion in All Dissertations by
an authorized administrator of TigerPrints. For more information, please contact [email protected].
 MODEL PREDICTIVE CONTROL OF MODERN HIGH-DEGREE-OF-FREEDOM
TURBOCHARGED SPARK IGNITED ENGINES WITH EXTERNAL COOLED EGR

A Dissertation
Presented to
the Graduate School
of Clemson University

In Partial Fulfillment
of the Requirements for the Degree
Doctor of Philosophy
Automotive Engineering

by
Rohit Vishvanath Koli
August 2018

Accepted by:
Dr. Robert Prucka, Committee Chair
Dr. Zoran Filipi
Dr. Qilun Zhu
Dr. Srikanth Pilla
 ABSTRACT

The efficiency of modern downsized SI engines has been significantly improved

using cooled Low-Pressure Exhaust Gas Recirculation, Turbocharging and Variable

Valve Timing actuation. Control of these sub-systems is challenging due to their inter-

dependence and the increased number of actuators associated with engine control. Much

research has been done on developing algorithms which improve the transient

turbocharged engine response without affecting fuel-economy. With the addition of

newer technologies like external cooled EGR the control complexity has increased

exponentially.

This research proposes a methodology to evaluate the ability of a Model

Predictive Controller to coordinate engine and air-path actuators simultaneously. A semi-

physical engine model has been developed and analyzed for non-linearity. The

computational burden of implementing this control law has been addressed by utilizing a

semi-physical engine system model and basic analytical differentiation. The resulting

linearization process requires less than 10% of the time required for widely used

numerical linearization approach. Based on this approach a Nonlinear MPC-Quadratic

Program has been formulated and solved with preliminary validation applied to a 1D

Engine model followed by implementation on an experimental rapid prototyping control

system.

The MPC based control demonstrates the ability to co-ordinate different engine

and air-path actuators simultaneously for torque-tracking with minimal constraint

violation. Avenues for further improvement have been identified and discussed.

ii
 DEDICATION

This dissertation is dedicated to my family in India. Without your support, this

endeavor would not have been possible for me.

iii
 ACKNOWLEDGMENTS

I am extremely grateful to my advisor Dr. Robert Prucka for his valuable

guidance and support. He is among the few academic advisors who proactively prioritize

the students’ well-being. He has taught me that the easiest person to convince is the

person in the mirror, which has inspired me to cultivate and maintain the rigor of

fundamental research.

I would also like to thank my friend Dr. Qilun Zhu for his continual input and the

several hour long discussions we have had regarding engine control, science and life. His

work has been a great source of inspiration for my research. Thank you for having the

patience and the will to help me to the best of your ability I would also like to thank Dr.

Zoran Filipi for his valuable advice on turbocharged engine control on modeling. His

class lecturers on advanced IC engine controls was a memorable experience for me.

The research work performed was in collaboration with industrial partners Robert

Bosch LLC and Fiat Chrysler Automobiles. I would like to thank Jason Schwanke and

Dr. Shyam Jade from Robert Bosch LLC for their valuable industrial insight and quick

assistance in this research. I would also like to thank Michael Prucka and the folks at Fiat

Chrysler Automobiles for their valuable inputs and collaboration.

Lastly, I would like to thank all my fellow CUICAR students and friends,

Konstantinos Siokos, Shuonan Xu, Harikesh Arunachalam, Anirudh Allam, Dan Egan,

Dennis Robertson, Tommy Powell, Ryan O’Donnell, and others for their help and

assistance.

iv
 TABLE OF CONTENTS

Page

TITLE PAGE .................................................................................................................... i

ABSTRACT ..................................................................................................................... ii

DEDICATION ................................................................................................................iii

ACKNOWLEDGMENTS .............................................................................................. iv

LIST OF TABLES .......................................................................................................... ix

LIST OF FIGURES ......................................................................................................... x

CHAPTER

I. INTRODUCTION ......................................................................................... 1

Down-sizing SI internal combustion engines .......................................... 1

v
Table of Contents (Continued) Page

Control problems associated with high-degree of freedom engines .......... 2

Motivation for utilization of Model Predictive Control ............................. 7

Challenges in implementing MPC on turbocharged SI engines ................ 8

Research objectives and outline ............................................................... 11

II. EXPERIMENTAL TESTING ARCHITECTURE ...................................... 13

Test Engine Setup .................................................................................. 12

Data Acquisition .................................................................................... 15

Rapid Control Prototyping Setup ........................................................... 16

III. CONTROL ORIENTED ENGINE MODELING ....................................... 20

Parametrized Turbocharger model ......................................................... 19

Data driven parameter models for engine cylinders .............................. 38

Continuous time state space model of the engine .................................. 47

IV. NON-LINEARITY ANALYSIS OF CONTROL ORIENTED MODEL ... 53

Local linear approximation .................................................................... 51

Step and sinusoidal input transient testing ............................................. 51

vi
Table of Contents (Continued) Page

V. NON-LINEAR MODEL PREDICTIVE CONTROL DESIGN .................. 62

Horizon length and sample time determination ..................................... 59

Non-linear system model linearization .................................................. 61

Model discretization in time .................................................................. 70

MPC Cost function formulation............................................................. 72

MPC Quadratic program problem formulation and optimization.......... 73

Co-simulation with GT-Power ............................................................... 74

VI. LOWER-LEVEL CONTROL AND ESTIMATION .................................. 84

Valve mass flow rate control ................................................................. 80

Low-pressure EGR ∆P control............................................................... 82

Turbo speed estimation and state measurement..................................... 82

VII. EXPERIMENTAL VALIDATION OF MPC ....................................... 89

Experimental control system hardware configuration ........................... 87

Experimental test results: Aggressive tip-in and tip-out maneuvers ..... 93

vii
Table of Contents (Continued) Page

VIII. CONCLUSIONS AND FUTURE WORK .......................................... 102

Summary of research work .................................................................... 97

Broader impacts and practical aspects ................................................... 99

Future work and potential for improvement ........................................ 100

APPENDICES ............................................................................................................. 107

A: A Control Algorithm for Low Pressure – EGR Systems using a Smith

Predictor with Intake Oxygen Sensor Feedback .................................. 102

B: Nonlinear Model Predictive Control of Dual Loop - Exhaust Gas

Recirculation in a Turbocharged Spark Ignited engine ....................... 125

REFERENCES ............................................................................................................ 152

viii
 LIST OF TABLES

Table Page
Table II-1. Engine specifications ...................................................................................... 14
Table III-1. Error metrics of the compressor mass flow rate models ............................... 25
Table III-2. Actuator combinations used for data collection for ANN model training .... 40
Table III-3 Inputs used for training ANNs ...................................................................... 42
Table III-4. Accuracy of trained ANNs ............................................................................ 48
Table III-5 Units used for normalization of state space model ......................................... 49
Table III-6 Error metrics of the 0D model for four different transient engine trajectories.
........................................................................................................................................... 51
Table IV-1 Minimum number of model linearizations required in horizon to maintain
IMEPn model accuracy for varying input frequency ........................................................ 57
Table IV-2 Minimum number of model linearizations required in horizon to maintain
IMEPn model accuracy for varying input frequency ........................................................ 59
Table A-1 Conventional PID control vs Smith Predictor Control performance comparison
......................................................................................................................................... 129
TABLE A-2. Performance summary of EGR % control using NMPC .......................... 150

ix
 LIST OF FIGURES

Figure Page

Figure I-1. High-degree of freedom SI Turbocharged engine ............................................ 2

Figure I-2. Load control using intake throttle and wastegate on downsized turbocharged
engines ................................................................................................................................ 4
Figure II-1.The General Motors LTG engine [92]............................................................ 13
Figure II-2. Exhaust side of the test engine ...................................................................... 15
Figure II-3. Intake side of tests engine.............................................................................. 15
Figure II-4. Engine and dynamometer interface ............................................................... 16
Figure II-5: The BOSCH MED17 Engine ECU ............................................................... 17
Figure II-6. Dspace rapid prototyping system .................................................................. 19
Figure II-7. ETAS ES910 rapid prototyping system ........................................................ 19
Figure III-1. Polynomial approximation for Jensen compressor model coefficients ........ 24
Figure III-2. Modeled compressor iso-speed lines vs manufacturer provided map points25
Figure III-3. Modeled compressor efficiency curve vs manufacturer provided compressor
map points ......................................................................................................................... 28
Figure III-4. 0D Modeled vs GT-Power modeled compressor inlet temperature ............. 29
Figure III-5 0D Modeled vs GT-Power modeled compressor outlet temperature ............ 30
Figure III-6. Modeled turbine mass flow rate vs manufacturer provided turbine map
points ................................................................................................................................. 32
Figure III-7. Turbine pressure ratio as a function of turbine mass flow rate with map
points for four different turbine inlet temperatures........................................................... 34
Figure III-8. Modeled turbine efficiency curve vs map points ......................................... 36
Figure III-9. Comparison of measured and GT-Power modeled transient turbocharger
speed response with tuned inertia for 1500 RPM ............................................................. 38
Figure III-10. Comparison of measured and GT-Power modeled transient turbocharger
speed response with tuned inertia for 2500 RPM ............................................................. 38
Figure III-11. Experimental setup to facilitate data acquisition for engine cylinder models
........................................................................................................................................... 41
Figure III-12. Modeled vs measured volumetric efficiency [%] for all data with +/- 10%
error bands represented by dashed red line ....................................................................... 43
Figure III-13. Modeled vs measured IMEPn for all data with +/- 10% error bands
represented by dashed red line .......................................................................................... 44
Figure III-14. Modeled vs measured 𝐶𝑂𝑉𝐼𝑀𝐸𝑃𝑛 for all data with +/- 10% error bands
represented by dashed red line .......................................................................................... 45
Figure III-15. Modeled vs measured 𝐾𝐼2 for all data with +/- 10% error bands
represented by dashed red line .......................................................................................... 46

x
Figure III-16. Modeled vs measured steady state exhaust temperature for all data +/- 10%
error bands represented by dashed red line. The accuracy of each of these neural
networks is summarized in................................................................................................ 47
Figure III-17 Open-loop system trajectory for 0D model vs experimentally collected data
........................................................................................................................................... 52
Figure IV-1 Deviation between the non-linear system model and the linearized system
model as a function of time step N from the linearization. The deviation is induced by
perturbing 𝑚𝐼𝑇𝑉 corresponding to commanded 𝐼𝑀𝐸𝑃𝑛 ................................................. 54
Figure IV-2 Model error as a function of commanded open-loop change in IMEPn
shows that the deviation between linearized and non-linear model varies as a function of
operating condition and the time –step after the point of linearization ............................ 57
Figure IV-3 Model error as a function of frequency of commanded open-loop model
input shows that the deviation between linearized and non-linear model generally peaks
between 1.5 Hz and 0.5 Hz depending on operating condition ..................................... 60
Figure V-1 Approximate relationship between horizon length considered for MPC, size
of the model and the computational expense of real-time execution................................ 62
Figure V-2 The selection of MPC sample time and prediction horizon time dictates the
choice of appropriate engine dynamics in the engine model ............................................ 63
Figure V-3. Graphical layout of Artificial Neural Network ............................................. 70
Figure V-4. Comparison between execution times of 𝐼𝑀𝐸𝑃𝑛 ANN model linearization
using numerical and analytical approach .......................................................................... 73
Figure V-5. Comparison between execution times of complete system model linearization
using numerical and analytical approach .......................................................................... 74
Figure V-6. Co-simulation setup between GT-Power and Simulink ................................ 79
Figure V-7 IMEPn response of MPC co-simulated with GT-Power ................................ 80
Figure V-8 Control actuator trajectories for MPC cosimulation with GT-Power, air-path
actuators are on the left hand side and cylinder actuators are on the right hand side. Black
dashed lines are the steady-state reference inputs and blue lines are the inputs generated
by MPC ............................................................................................................................. 81
Figure V-9 Compressor operation during co-simulation of GT-Power with MPC .......... 83
Figure VI-1. The gradient of the flow function becomes very steep as the pressure
difference across the valve diminishes ............................................................................. 85
Figure VI-2.Real-time estimated vs measured turbocharger speed using fixed point
iteration method ................................................................................................................ 88
Figure VII-1. Rapid control prototyping setup for experimental validation of MPC ....... 90
Figure VII-2. Tip-in IMEPn response for MPC and Open-loop control strategy............. 91
Figure VII-3. Intake throttle mass flow rate trajectories for MPC and Open-loop control
........................................................................................................................................... 92
Figure VII-4. Wastegate mass flow rate trajectories for MPC and Open-loop control .... 92
Figure VII-5. External cooled EGR mass flow rate trajectories for MPC and Open-loop
control shows that MPC deliberately stops EGR flow on the aggressive tip-in ............... 93
Figure VII-6. Spark timing trajectories for MPC and Open-loop control ........................ 94
Figure VII-7. Camshaft position timing trajectories for MPC and Open-loop control .... 94

xi
Figure VII-8 ANN Modeled volumetric efficiency using MPC inputs and open-loop
inputs for tip-in maneuver ................................................................................................ 95
Figure VII-9 Measured turbine inlet temperature for MPC and Open-loop control shows
that MPC commands late spark timing to increase specific exhaust gas enthalpy for
turbocharger response ....................................................................................................... 95
Figure VII-10 Turbocharger speed for MPC increases much more rapidly than open-loop
control ............................................................................................................................... 96
Figure VII-11 Measured 𝐾𝐼2 for MPC does not exceed the limit of 1 despite advanced
spark timing during tip-in ................................................................................................. 97
Figure VII-12. Tip-out IMEPn response for MPC and Open-loop control strategy......... 98
Figure VII-13 Air-path trajetories for aggressive tip-out maneuver ............................... 100
Figure VII-14 Cylinder actuator trajectories for aggressive tip-out maneuver............... 100
Figure VII-15 Compressor operation during the aggressive tip-out ............................... 101
Figure A-1 Control system setup .................................................................................... 110
Figure A-2 Dimensional Discharge coefficient map obtained from optimization routine
......................................................................................................................................... 113
Figure A-3 Measured vs modeled EGR valve mass flow rate ........................................ 114
Figure A-4 Modeled low pressure EGR path ................................................................. 115
Figure A-5. Closed loop EGR fraction controller with Smith predictor......................... 118
Figure A-6 Modeled vs measured transport delay .......................................................... 120
Figure A-7. Comparison between modeled and measured EGR fraction at EGR sensor
location; engine operated at 1700 RPM 3 Bar BMEP .................................................... 121
Figure A-8. Comparison between modeled and measured EGR fraction at EGR sensor
location; engine operated at 2100 RPM 5 Bar BMEP .................................................... 122
Figure A-9. Comparison between modeled and measured EGR fraction at EGR sensor
location; engine operated at 2100 RPM 2.5 Bar BMEP ................................................. 123
Figure A-10. Step response of EGR fraction controller with conventional PID controller
at 1500RPM 7 Bar BMEP .............................................................................................. 124
Figure A-11. Step response of EGR fraction controller with Smith Predictor at 1500RPM
7 Bar BMEP .................................................................................................................... 125
Figure A-12. Step response of EGR fraction controller with conventional PID controller
at 1300RPM 4 Bar BMEP .............................................................................................. 126
Figure A-13. Step response of EGR fraction controller with Smith Predictor ............... 127
Figure A-14. Step response of EGR fraction controller with conventional PID controller
at 2000RPM 9 Bar BMEP .............................................................................................. 128
Figure A-15. Step response of EGR fraction controller with Smith Predictor at 2000RPM
9 Bar BMEP .................................................................................................................... 128
Figure A-16. Schematic of Engine and Air path model.................................................. 133
Figure A-17. Step response of oxygen sensor based EGR measurement installed at
compressor outlet location .............................................................................................. 135
Figure A-18. Comparison of step response of transport delay models with true dead time
delay response ................................................................................................................. 136
Figure A-19 Closed-loop control structure of DL-EGR NMPC..................................... 139

xii
Figure A-20 Pressure states and intake manifold EGR dilution trajectories for HP-EGR
control only. .................................................................................................................... 141
Figure A-21 Input trajectories for HP-EGR control only. .............................................. 142
Figure A-22 Pressure states and intake manifold EGR dilution trajectories for LP-EGR
control only. .................................................................................................................... 144
Figure A-23 Input trajectories for LP-EGR control only................................................ 145
Figure A-24 Pressure states and intake manifold EGR dilution trajectories for Dual-loop
EGR control. ................................................................................................................... 147
Figure A-25 Input trajectories for Dual-loop EGR control. ........................................... 148
Figure A-26 Transient intake manifold EGR dilution for multiple EGR architectures .. 149

xiii
 I. INTRODUCTION

1.1 Down-sizing SI internal combustion engines

The practice of reducing engine displacement whilst maintaining and surpassing

performance requirements using turbocharging i.e. ‘downsizing’ has been widely

followed by the Automotive industry [1][2][3][15]. The recently projected industrywide

requirement of a fleet average of 54.5mpg for the model year 2025 requires radical

improvements in fuel economy [4]. This can be facilitated partially by further downsizing

with more sophisticated methods of turbocharging. Classically, the performance of

turbocharged engines during transient engine operation is influenced strongly by the

turbocharger behavior. Since turbochargers are mechanically independent from the

engine, there are undesired transient phenomenon which severely affect performance and

consequently driver perception during these situations. This phenomenon is called the

‘turbo-lag’ and is literally a time delay between the driver-demanded engine torque and

the response of the engine. Dedicated systems have been widely researched and utilized

to improve turbocharger response [5][8][9][10]. However, these systems usually result in

transient fuel economy and emissions penalties.

Lighter turbocharger rotor assemblies have also been investigated by [6][7] which

result in quicker turbocharger acceleration due to lower moment of inertia. Utilization of

Variable Geometry Turbines (VGT) has also shown significant improvement in transient

response on turbocharged diesel applications [19][22]. VNT application on SI engines

has not only shown improvement in transient response but also improved low speed

1
boosting ability [23]-[26]. However, exhaust gas temperature, cost, and reliability

limitations have impeded penetration of this technology into mainstream gasoline

powered vehicles. Electrical turbocharger compounding concepts have been studied in

[26][28] for improved turbocharger response and exhaust heat recovery. All of these

solutions which focus on improved hardware characteristics and performance result in

elevated cost and added system complexity which is unfavorable for mass production.

1.2 Control problems associated with high-degree of freedom engines

Figure I-1. High-degree of freedom SI Turbocharged engine

Modern engines are highly over-actuated with respect to torque control, wherein

multiple actuator combinations exist for a single engine torque level at a given engine

2
speed. Usually a unique combination exists which optimizes the steady-state fuel

economy of the engine. However, multiple optimum combinations with negligible fuel

economy difference may exist [11][12]. The actuator combinations themselves can be

very different from each other which further confounds the determination of the

appropriate steady state engine-actuator combination.

The dynamics of the different subsystems on the engine are also vastly different

in timescales. The impact of each actuator on its corresponding subsystem and coupling

with other subsystems affects the overall engine’s dynamic response drastically. Because

of the fluid coupling control of the turbocharger actuators is critical for engine

performance and fuel economy. Early turbocharger control methods have been outlined

[14] with introduction to rule based electronic control of waste-gate and blow-off valve.

One of the most significant challenges associated with turbocharged engine control has

been the coordination of the waste-gate and the intake throttle valve in turbocharged SI

engines. The primary load control actuator for SI engines is the intake throttle valve.

However, in order to increase the intake manifold pressure beyond atmospheric pressure,

the turbocharger is utilized. Exhaust enthalpy drop across the turbine results in power

transfer to the compressor. The compressor performs work on the inlet air, thereby

increasing it’s temperature and pressure, also known as ‘boost’ pressure. The waste-gate

valve, which controls the turbine bypass flow has been typically used to limit the

maximum boost pressure using a pre-calibrated mechanical waste-gate actuator.

However, with the advent of electronic waste-gate control this valve is kept wide open

under throttled operation to minimize the exhaust pressure and hence minimize pumping

3
work performed by the engine. However, under transient operation, the control of the

intake throttle and waste-gate valve requires complex strategies to simultaneously

minimize fuel consumption and improve turbocharger response.

Figure I-2. Load control using intake throttle and wastegate on downsized turbocharged

engines

Much research has been conducted in the control and coordination of the

turbocharger and air-path actuators since then with an exhaustive list of papers and

patents [13][17]. Karnik, et. al have evaluated the performance of decentralized Single

Input Single Output (SISO) controllers vs full-state feedback and the reduced versions of

full state feedback control in [16] using linearized engine models. The results of this

work begin to show the benefit of utilizing Multi-Input-Multi-Output (MIMO) control

laws for waste-gate and intake throttle coordination. Typically each of the subsystems

4
have been controlled with individual control loops targeting a particular set-point as

outlined in [13]. Non-linear control approaches have been introduced in [17] using

Parallel Distributed Compensation control. Turbocharger waste-gate control under

boosted conditions is explored in [18], where singular perturbation is utilized to obtain a

first order non-linear model. The feed-forward control law is derived based on inversion

of the reduced order model, which is then augmented with an error-integral based

feedback term. The potential of Variable Valve Timing (VVT) to improve turbocharger

response has been mentioned in [59][60] by improving air-flow into the cylinders and in

some conditions directly from the intake port to the exhaust port during the valve overlap

phase. Colin et al. [57][58] have proposed a decentralized control system which consists

of feed-forward intake throttle controller and NMPC based waste-gate controller for

engine torque control. The control of Variable Valve Timing (VVT) into the engine

torque controller for control of the trapped Residual Gas Fraction (RGF) has been

introduced using a static neural network model. Interaction between the turbocharger

response and VVT has been discussed but not exploited in the control strategy due to the

decentralized structure of the controller which targets RGF and Torque independently.

Spark timing also has a significant impact on the combustion process, exhaust

temperature and the cycle to cycle variability of combustion [54][71]. For best fuel-

economy, the spark-timing is calibrated to position fifty percent of the in cylinder fuel–air

mixture burnt at eight crank angle degrees after top dead center. However, on

turbocharged SI engines the spark timing at higher loads is often limited by the onset of

knocking in the cylinder due to end-gas auto-ignition [65] [68]. Such abnormal

5
combustion causes premature damage to the engine. The Spark timing is retarded to a

safe level such that knocking is minimized to acceptable levels. This is called the knock-

limited-spark-advance timing. Retarding spark timing for temporary increase of the

specific enthalpy of the exhaust gas have been shown to reduce turbocharger lag in

[72]and [9]. There is significant potential to improve transient response by utilizing this

actuator. However, due to the knock and combustion stability constraints mentioned

above, simultaneous control of spark timing along with other engine actuators is

challenging.

Current generation production systems generally comprise multiple Single Input

Single Output (SISO) type controllers with lookup-tables and maps. Although this

approach is simple to adopt for OEMs, it requires significant calibration effort and tuning

to ensure stability and robustness. This entails significant real-time testing on the

dynamometer and in some cases the vehicle as well. The recent findings of implementing

Exhaust Gas Recirculation (EGR) in SI turbocharged engines has also shown potential to

improve fuel economy drastically [11][12][29]-[36]. However, this adds yet another

control actuator which needs to be considered in the overall engine control system.

Moreover, low-pressure cooled EGR also has known issues associated with the inability

of feed-forward flow modeling through the valve and transport delay of the EGR Air

mixture through the intake air-path [37]-[41]. This delay is more significant for SI

engines compared to CI engines because the combustion process of SI engines is far more

sensitive to external dilution, since external EGR has a significant effect on the

combustion stability and knock propensity.

6
 Because the number of ‘control knobs’ on engines is ever increasing OEMs are

beginning to realize the potential of using model based control methodologies which

could significantly reduce the required calibration and tuning efforts. Multi-Input-Multi-

Output (MIMO) control systems have shown significant promise in being capable of

coordinating multiple engine actuators based on multiple targeted outputs. One particular

MIMO control approach which has gained attention in recent years is Model Predictive

Control (MPC).

1.3 Motivation for utilization of Model Predictive Control

Model Predictive Control is a class of control algorithms in which the control

actions for the system are generated as a result of an optimization problem over a fixed

time period into the future from the current time. This period is called the prediction

horizon. As the closed loop control system executes and proceeds in time, the horizon

recedes further equally. MPC is hence also known as Receding Horizon Control (RHC)

This optimization problem is formulated from a mathematical objective function, the

current condition and a dynamic control oriented model of the system. The general MPC

algorithm can be summarized as follows:

1. Generate a predicted future reference trajectory of the system for the prediction

horizon based on a known model, current state of the system and reference inputs.

2. Formulate and solve an optimization problem based on a user-defined objective

function related to the behavior of the system for the prediction horizon.

3. Apply the inputs that correspond to the current time to the system

7
 4. Re-evaluate the state of the system at periodic intervals and repeat the process

It is clearly visible that this strategy automatically considers the slow and the fast

dynamics of the system simultaneously due to pre-evaluated knowledge of the system’s

behavior for a specific time period in the future. The main advantage of applying MPC to

engines is the significant reduction of calibration and tuning efforts required. MPC tuning

can be achieved easily by simple manipulation of the constant terms within the objective

function. This is greatly simplified compared to the conventional tuning and calibration

approach used for conventional SISO and LUT based methods. Another major advantage

of using MPC is the ability to handle constraints. These constraints can be simple

saturation limit constraints for the control actions or more complex operational

constraints. This attribute is particularly favorable for turbocharged engines. Operational

constraints for turbocharger protection, combustion stability and knocking which are

extremely detrimental to the mechanical integrity, fuel economy and exhaust emissions

can be efficiently incorporated into the MPC framework.

1.4 Challenges in implementing MPC on turbocharged SI engines

The predominant challenge in widespread implementation of MPC in production

is the real-time computational ability. In order to execute MPC in real-time a sufficiently

powerful microcontroller with ample memory is required. In order to justify the usage of

MPC for engine control in production, the MPC has to be capable of operating at any

engine operating condition and any kind of engine transient, with consideration of fuel

8
economy, emissions, hardware protection, drivability and most importantly torque

delivery. The dynamic engine model complexity and accuracy begins to become a

concern when all of the above mentioned aspects have to be considered.

Internal combustion engine behavior is highly complicated and requires high-

order models to capture high levels of accuracy and mimic relevant dynamic behavior.

The addition of a turbocharger to the engine introduces additional coupled, non-minimum

phase dynamics which complicates the model further. The turbocharger’s inertial

dynamics have a time constant which is in the order of seconds. This is significantly

slower compared to the cycle by cycle dynamics of the engine cylinders. Incorporating all

of these dynamics into one prediction model and consideration of a time period long

enough to account for the turbocharger’s inertial dynamics results in MPC based

optimization problems which are impractical to execute in real-time even on research-

grade high performance micro-controllers. Hence, choosing an appropriate prediction

model for the MPC is the most important task.

A linear prediction model is the favorable choice for MPC formulation due to it’s

simplicity in implementation. Using a fixed pre-determined linear model also eliminates

the computational cost of obtaining a linearized model from a more accurate non-linear

system model in real-time. The prediction model utilized in many of the past research

articles is a local linear approximation of the engine dynamics [73][75]. Many of these

research articles focus mainly on the control of the air-path of the diesel engines using

predetermined equilibrium-point linearized system models stored in the memory of the

micro-controller. Some of these articles utilize black-box models identified directly from

9
experimental or high-fidelity simulation data of the engine. The number of models used

to approximate the engine dynamics is also an open question since utilizing lesser models

results in larger inaccuracy in the prediction of system behavior during transients and

complicates the transitioning between linear models. These approaches have been

demonstrated to work well mainly because of the linear-like behavior of the air-path

dynamics (intake manifold and compressor outlet pressure). However, this approach has

not been demonstrated to track highly non-linear outputs associated with engine behavior

like engine torque while simultaneously adhering to combustion constraints which are

also complicated non-linear functions of the system states, inputs and parameters.

Application of MPC to gasoline SI engines has been recently explored by Wiese

et al in [77] for evacuation of EGR trapped in the intake manifold using a Linear Time

Varying MPC. This approach demonstrates the ability of the MPC to schedule the VVT

actuation in order to short-circuit the flow from intake ports directly to the exhaust port

using available preview of the future torque reference. Bemporad et al have explored the

ability of utilizing this short-circuit flow in order to improve turbocharger boost response

in [82] using a linear MPC with pre-stored linear black-box models to control the air-path

actuators. Similar to the diesel MPC research, the air-path dynamics of the SI

turbocharged engine are also linear-like, and hence the local-linear formulation has been

shown to work.

Naturally aspirated SI engine torque control while simultaneously adhering to

actuator and combustion constraints has been recently investigated in [79]-[81],[87] and

[54]. Combustion variability and knock constraint have been considered in the

10
formulation. Sequential Quadratic Programming based approach is used to solve the

MPC iteratively. The same approach has been further investigated in [83] for Air-path

control of a SI turbocharged engine which has a continuous surge control valve.

1.5 Research objectives and outline

This research focuses on developing an MPC framework using a control oriented

engine model of a SI turbocharged gasoline engine. The air-path model is largely

physical and the cylinder model is purely data driven. The goal is to unify the control of

air-path and engine cylinder actuators into a single control system which automatically

co-ordinates these actuators to improve transient turbocharged engine response while

respecting the engine’s operational constraints. In order to achieve this goal many sub-

tasks are required to be performed.

1. Develop a control oriented model of the engine which captures the necessary and

sufficient dynamic behavior of the engine

2. Develop an algorithm that utilizes the control oriented model and formulates an

MPC based optimization problem with a user defined objective function

comprising of multiple reference targets

3. Identify avenues and implement strategies to minimize the computational burden

of formulating the MPC based optimization problem

4. Solve the MPC based optimization problem by utilizing a solver algorithm

capable of executing in real-time on a research grade rapid-prototpying micro-

controller

11
5. Build a Simulation framework to tune the MPC controller by coupling it with a

virtual engine developed using a 1D-Engine simulation software.

6. Develop an experimental validation framework for real-time testing of the MPC

12
 II. EXPERIMENTAL METHODS AND RESEARCH ENGINE SETUP

2.1 Test engine description

The experimental test engine is a turbocharged spark-ignited gasoline engine

designed by General Motors. It is codenamed as the GM-LTG engine and is available in

several standard and premium production vehicles from General Motors.

Figure II-1.The General Motors LTG engine [92]

The LTG engine is a representative of modern downsized engines with multiple

actuators. Additionally a low-pressure, cooled EGR circuit was added to the stock engine

as well. A detailed list of engine specifications is outlined in Table II-1

13
Table II-1. Engine specifications

Displacement 2.0L

Compression ratio 9.5:1

Air-EGR induction system Turbocharged with Low-pressure EGR

Fuel System Gasoline – Direct Injection

Bore 86

Stroke 86

Valve train DOHC 4 valves per cylinder

Maximum engine speed 7000 RPM

An aftermarket stainless-steel EGR cooler for a Ford F250 is utilized as the LP-

EGR system. The EGR valve from the Audi 1.8TDI is installed as the LP-EGR flow

control valve. Engine coolant is circulated through the LP-EGR cooler to emulate real-

world cooled EGR systems used in production passenger car diesel engines. Excessive

EGR cooling has also been known to cause condensation in the compressor inlet location

leading to water droplet impingement on the compressor vanes. An additional pre-

compressor throttle is also installed upstream of the EGR mixing location. The original

pneumatic waste-gate actuator has been replaced with an electronic linear actuator for

improved position control of the waste-gate valve. Figure II-3, Figure II-2 show the hot

side as well as the cold side of the engine. The actuators and systems most relevant to

turbocharged SI engine control are shown in these figures.

14
Figure II-2. Exhaust side of the test engine

Figure II-3. Intake side of tests engine

15
2.2 Data acquisition and dynamometer interface

High and low speed data acquisition systems have been utilized on the test engine. The

high speed data acquisition consists of the crank angle resolved pressure measurement of

the cylinders, intake and exhaust ports and across the low pressure EGR valve. The low

speed data acquisition systems mainly consists of time averaged pressure, temperature

and mass flow rate measurements at various locations on the engine. The test engine

crankshaft is connected to the dynamometer using a custom flywheel and driveshaft.

FEV’s Test automation system is utilized to control the speed and torque setpoint of the

dynamometer.

Figure II-4. Engine and dynamometer interface

16
2.3 Rapid control prototyping hardware

Since the base engine is a production engine, it has production intent sensors and

actuators associated with basic engine functionality. An open Bosch ECU MED17 has

been interfaced with the production engine sensors and actuators. This ECU is also

capable of real time cylinder pressure feedback based diagnostics. The embedded

application software deployed onto the ECU has CCP bypass functionality which allows

run-time read-access to all of the ECU variables and write-access to the variables

associated with spark timing, camshaft timing and throttle position.

Figure II-5: The BOSCH MED17 Engine ECU

17
Additional subsystem like continuous blow off valve, electronic wastegate control

actuator and pre-compressor throttle are controlled using the ETAS ES930 and ES910

units. The ES910 is a rapid control prototyping ECU and calibration interface for the

Bosch MED17 ECU. The control applications are initially developed in

Simulink/Stateflow and compiled using the Intecrio Realtime system target file. This

target file compiles the base level Simulink/Stateflow model into a Scoop IX file with the

extension .six. This file is then imported into a software called ETAS Intecrio which

facilitates interfacing with the MED17 ECU, ES930 and other systems via CAN. The

execution rate and several other diagnostics can also be configured in Intecrio. The

ES910’s real-time computational ability is ‘limited’ with respect to running complicated

algorithms with large memory requirement and/or small execution time step. These

algorithms include computationally intensive Model Predictive Control, Moving Horizon

Estimation and complex plant-models for Hardware In the Loop simulation. For such

cases the Dspace DS1006 is utilized. The DS1006 system used has four processor boards

which can execute different applications independently. This system has high speed CAN

controllers which can be interfaced with other hardware. One of these controllers is

utilized to establish communication with the ES910, while the other is utilized to

communicate with the dynamometer control system for vehicle level simulations with

Engine In the Loop capability.

18
Figure II-6. Dspace rapid prototyping system

Figure II-7. ETAS ES910 rapid prototyping system

19
 III. CONTROL ORIENTED ENGINE MODELING

A modular approach has been adopted for modeling, wherein the individual

components of the air path have been modeled separately and ultimately combined

together to form a control oriented air path model.

3.1 Parametrized Turbocharger model

Selection of the turbocharger model is critical for model accuracy and execution

speed of the engine model. The turbocharger models are split into the compressor models

and the turbine models.

3.1.1 Compressor mass flow rate models

To model the pressure dynamics in the boost manifold the compressor mass flow

rate is modeled as a function of turbocharger speed and the pressure ratio across the

compressor. There are several models available in literature with varying degrees of

accuracy and targeted applications. Artificial Neural Network (ANN) based models have

been proposed in [47]. Although the prediction of these models within the compressor-

map is good, low speed extrapolation is an issue with this approach. Furthermore, the

structure of the ANN required to ensure in-map and low-speed extrapolation accuracy

changes as a function of the specific compressor chosen as mentioned in [48]. Hence, a

physics based classical approach was chosen to model the compressor mass flow rate.

Two models were evaluated for the compressor used in this research. The first one is the

ellipse model [44][46] The ellipse model is based on the ellipsoidal parametrization of

the iso-speed lines obtained from steady state gas-stand tests of the compressor. These

20
results of these tests are called the compressor maps and are provided by the

manufacturer for normal operating region of the compressor. The ellipse model supports

compressor mass flow rate prediction in the mild and deep surge region as well as the

choke region. However, the disadvantage of this approach is that the model changes

mathematical structure when transitioning from the normal region to the surge region of

the compressor map. Incorporating a variable structure model into MPC based control

algorithms is not favorable. Hence only the normal region of the compressor map was

modeled using the Ellipse model. The compressor map is modeled in the normal region

as follows.
1
Π 𝐶2 𝐶1
𝑚̇𝐶𝑜𝑚𝑝 = (𝑚̇𝑀𝑎𝑥 − 𝑚̇𝑍𝑆𝐿 ) [1 − (Π ) ] + 𝑚̇𝑍𝑆𝐿 1
ZSL

𝐶1 = 𝑓𝐶 1 (𝜔𝑇 ) = 𝐶1,0 + 𝐶1,1 𝜔 𝑇 2

𝐶
𝐶2 = 𝑓𝐶 2 (𝜔 𝑇 ) = 𝐶2,0 + 𝐶2,1 𝜔𝑇2,2 3

𝑚̇𝑀𝑎𝑥 = 𝑓𝑚̇𝑀𝑎𝑥 (𝜔 𝑇 ) = 𝐶3,0 + 𝐶3,1 𝜔 𝑇 4

𝑚̇𝑍𝑆𝐿 = 𝑓𝑚̇𝑍𝑆𝐿 (𝜔 𝑇 ) = 𝐶4,1 𝜔 𝑇 5

𝐶
ΠZSL = 𝑓𝑍𝑆𝐿 (𝜔 𝑇 ) = 1 + 𝐶5,1 𝜔𝑇5,2 6

Where, 𝑚̇ is the mass flow rate and 𝐶𝑖,𝑗 have real values. Π is the pressure ratio

across the compressor. 𝜔 𝑇 is the turbocharger’s rotational speed. Within the

manufacturer’s compressor map, 𝐶1 and 𝐶2 are identified by utilizing global, least-

21
squares optimization to minimize the difference between the manufacturer’s and modeled

iso-speed lines. The suffixes 𝑍𝑆𝐿 and 𝑀𝑎𝑥 correspond to two distinct locations on the

iso-speed line. The former corresponds to the point on the surge line and the latter

corresponds to the point on the Π = 0 line.

The second compressor model is proposed by Hadef et al. [49][50] which is an

evolution of the Jensen and Kristensen empirical mean value model [51]. The latter has

been widely used in publications relating to modeling, control and estimation of air path

on turbocharged engines. At the heart of these models is the normalization of the

nonlinear relationship between the compressor mass flow rate and the pressure ratio using

a dimensionless head parameter 𝜓 and dimensionless flow parameter 𝜙. These

parameters are defined as follows:

𝐾−1
[𝐶𝑃𝐶𝑖𝑛 𝑇𝐶𝑖𝑛 (Π 𝐾 −1)]
𝜓 = 𝜓(Π, 𝜔 𝑇 ) = 7
0.5𝑈𝐶2

𝑚̇𝐶𝑜𝑚𝑝 𝑅𝑇𝐶𝑖𝑛
𝜙= (𝑃𝐶𝑖𝑛 𝑑𝑐3 𝜔𝑇 )
8

Where, 𝐶𝑃𝐶𝑖𝑛 and 𝑇𝐶𝑖𝑛 are the constant pressure specific heat capacity and

temperature of the gas at the inlet of the compressor. 𝑅 and 𝐾 are the specific gas

constant and the ratio of specific heat capacity respectively. 𝑈𝐶 is the compressor blade

tip speed which is calculated as a function of the turbocharger speed 𝜔 𝑇 and the

compressor wheel diameter 𝑑𝑐 as follows.

𝜋
𝑈𝐶 = 60 𝑑𝑐 𝜔 𝑇 9

22
 The flow parameter 𝜙 is not directly measureable and in fact it is the parameter

which needs to be predicted in order to derive the compressor mass flow rate. Several

functions have been explored to derive 𝜙 from 𝜓 depending on the shape of the iso-speed

lines of the compressor [52][53]. The most widely used relationship is the inverse

proportionality function which uses the Mach number across the ring orifice of the

compressor. However, it requires additional determination of coefficients which are

derived from the turbo speed dependent Mach number. Preserving the structure of the

function 𝜙 is modeled as a function of the head parameter as follows as described in

[49][50].

𝐶(𝜔𝑇 )−𝐴(𝜔𝑇 )𝜓
𝜙= 10
𝐵(𝜔𝑇 )+𝜓

Where, 𝐴, 𝐵, and 𝐶 are modeled as second order polynomial functions of the

turbocharger speed. The coefficients of these polynomial functions are obtained similar to

the method used for obtaining 𝐶1 and 𝐶2 for the ellipse model.

23
 Figure III-1. Polynomial approximation for Jensen compressor model coefficients

The modeled iso-speed lines of both of the models are shown in Figure III-3. The

accuracy metrics for predicted mass flow rate as a function of compressor pressure ratio

for both the models is shown in Table III-1. Error metrics of the compressor mass flow

rate models. Although the Ellipse model accuracy is superior to the Jensen model, the

ellipse model requires a different mathematical structure in the surge region.

Determination of whether the compressor is operating in surge region is not possible

without the prior knowledge of the mass flow rate across it. The Jensen model structure

24
can be utilized all the way till the zero compressor mass flow rate. However, the true

compressor behavior in the deep surge region is not an accurate representation of the

compressor behavior predicted by the Jensen model. Since the scope of this research is to

not model these abnormal modes of operation, the Jensen model is used since it

transitions continuously from the normal operation region to the surge region

Figure III-2. Modeled compressor iso-speed lines vs manufacturer provided map points

Table III-1. Error metrics of the compressor mass flow rate models

Model RMS error [kg/min] Max error [kg/min]

Ellipse model 0.04 0.1

Jensen model 0.09 1.14

25
The compressor model is a quasi-steady state model [49]. During highly transient engine

operation Π can be calculated to have a value which is much higher than the maximum

possible pressure ratio that the compressor model can deliver at the corresponding

turbocharger speed and zero mass flow rate. This leads to an erroneous computation of 𝜓

and hence, to limit the operation of the compressor model to have realistic 𝜙 values even

during the transients, the 𝜓 parameter is saturated as follows.

𝐶(𝜔 )
𝜓𝑀𝑎𝑥 = 𝐴(𝜔𝑇 ) 11
𝑇

𝜓𝐹𝑖𝑛𝑎𝑙 = min⁡(𝜓𝑀𝑎𝑥 , 𝜓) 12

3.1.2 Compressor efficiency model

The compressor efficiency can vary significantly depending upon the operating

region on the compressor map. Assumption of a constant compressor efficiency has a

significant impact with regards to turbocharger inertial dynamics since the turbocharger

speed dynamics are directly coupled with this parameter. The compressor isentropic

efficiency 𝜂𝐶𝑜𝑚𝑝 is defined using the following equation.

𝑃
𝜂𝐶𝑜𝑚𝑝 = 𝑃 𝐶⁡𝐼𝑑𝑒𝑎𝑙 13
𝐶⁡𝐴𝑐𝑡𝑢𝑎𝑙

Where, 𝑃𝐶⁡𝐼𝑑𝑒𝑎𝑙 is the power required to compress the gas from inlet to the outlet

pressure, defined by the isentropic process. 𝑃𝐶⁡𝐴𝑐𝑡𝑢𝑎𝑙 is the actual power required by the

compressor to pressurize the gas. 𝑃𝐶⁡𝐴𝑐𝑡𝑢𝑎𝑙 is always lesser than 𝑃𝐶⁡𝐼𝑑𝑒𝑎𝑙 due to the heat

26
transfer and other losses associated with the pumping of inlet gas. Since 𝑃𝐶⁡𝐴𝑐𝑡𝑢𝑎𝑙 is not

measured on the engine, determination of 𝜂𝐶𝑜𝑚𝑝 cannot be made directly from this

equation. [53] has shown that it can be identified as a function of 𝜙. However, 𝜙 is the

parameter which is derived from 𝜓. Hence, it is identified directly as a polynomial

function of the 𝜓 as follows.

𝜂𝐶𝑜𝑚𝑝 = 𝑏1 𝜓3 ⁡ + ⁡ 𝑏2 𝜓 2 + ⁡ 𝑏3 𝜓⁡ + ⁡ 𝑏4 14

Where, 𝑏1 , 𝑏2 , 𝑏3 and 𝑏4 are identified to minimize the error between the manufacturer’s

data and the modeled 𝜂𝐶𝑜𝑚𝑝 . Figure III-3 shows the fit between the modeled and the

manufacturer’s compressor efficiency data. Majority of the modeled compressor

efficiency curve is within 5% of the manufacturer’s data with some points exceeding the

5% error band marginally.

27
Figure III-3. Modeled compressor efficiency curve vs manufacturer provided compressor

map points

3.1.3 Compressor outlet temperature model

The compressor outlet temperature is simply calculated from the computed

compressor efficiency and the pressure ratio across the compressor as follows.

𝐾−1
𝑃 𝐾
( 𝐵𝑠𝑡 ) −1
𝑃𝐶𝑖𝑛
𝑇𝐶𝑜𝑚𝑝𝑂𝑢𝑡 = 𝑇𝐶𝑖𝑛 ( + 1) 15
𝜂𝐶𝑜𝑚𝑝

Where, 𝑇𝐶𝑖𝑛 is the compressor inlet temperature calculated using the adiabatic mixing

equation as follows.

28
 𝑚̇𝐸𝑔𝑟 𝑇𝐸𝑔𝑟 +𝑚̇𝐴𝑖𝑟 𝑇𝐴𝑖𝑟
𝑇𝐶𝑖𝑛 = 16
𝑚̇𝐸𝑔𝑟 +𝑚̇𝐴𝑖𝑟

Where, 𝑚̇𝐸𝑔𝑟 is assumed to be modeled and 𝑚̇𝐴𝑖𝑟 is directly measured from the

engine air mass flow rate sensor which is a few inches upstream of the EGR mixing

location. 𝑇𝐸𝑔𝑟 is the temperature of the EGR and it is assumed to be available as a

measurement. Preliminary validation against GT-Power simulation of the compressor

inlet and outlet temperature models is shown in Figure III-4 and Figure III-5

Figure III-4. 0D Modeled vs GT-Power modeled compressor inlet temperature

29
Figure III-5 0D Modeled vs GT-Power modeled compressor outlet temperature

3.1.3 Turbine mass flow rate and pressure ratio model

Two turbine mass flow rate models have been developed. One model which

predicts the mass flow rate across the turbine as a function of the pressure ratio across it

and another model which predicts the pressure ratio as a function of known mass flow

rate across the turbine. The turbine mass flow rate is represented as a reduced Turbine

Flow Parameter as follows.

30
 𝑚̇𝑇𝑢𝑟𝑏 √𝑇𝑇𝑖𝑛 105
𝑇𝐹𝑃 = 17
𝑃𝑇𝑖𝑛

The manufacturer’s turbine data is represented in this format and hence

conversion to mass flow rate is required in order to fit a model to the data. The model

used to predict turbine mass flow rate as a function of pressure ratio is a modified version

of the standard restriction model described by Eriksson in [10] has been utilized to model

the relationship between the pressure ratio and mass flow rate across the turbine. The

structure of this model is very similar to the simplified orifice flow model and is defined

as follows.

𝐴𝑇𝑢𝑟𝑏 .P𝑇𝑖𝑛 𝑃
𝑚̇ 𝑇𝑢𝑟𝑏 = . ψPR (√𝑃 𝑇𝑖𝑛 ) 18
√𝑅. 𝑇𝑇𝑖𝑛 𝑇𝑜𝑢𝑡

Where, the flow function ψPR is given by,

1
𝑃𝑇𝑖𝑛 𝑃𝑇𝑖𝑛 𝑃𝑇𝑖𝑛 2
ψPR (√𝑃 ) = (2√𝑃 [1 − √𝑃 ]) 19
𝑇𝑜𝑢𝑡 𝑇𝑜𝑢𝑡 𝑇𝑜𝑢𝑡

The area parameter 𝐴𝑇𝑢𝑟𝑏 is modeled as a linear function of turbocharger speed

𝐴𝑇𝑢𝑟𝑏 = 𝑗1 𝜔𝑇 + 𝑗2 20

The parameters 𝑗1 and 𝑗2 are identified using the manufacturer’s data. The main

difference between this flow function and the one used in the orifice model is that ψPR is

applied to the square root of the pressure ratio across the turbine as mentioned in [53].

The flow function of the standard orifice model changes its structure when the pressure

31
ratio reduces below the critical pressure ratio. This causes the standard orifice model to

operate under the choked flow regime where the flow function ψPR loses coupling with

the turbine outlet pressure. Utilizing the square root of the pressure ratio instead of the

pressure ratio itself, significantly improves the accuracy of the restriction based model.

Figure III-6 shows the modeled and the manufacturer’s turbine TFP points.

Figure III-6. Modeled turbine mass flow rate vs manufacturer provided turbine map

points

Due to it’s non-linear nature and the square root terms present in the model, the

inversion of this restriction based model is very difficult. Due to this concern, a simple 3rd

order polynomial model is defined which predicts turbine pressure ratio as a function of

mass flow rate as follows.

32
𝑃𝑇𝑖𝑛
𝑃𝑇𝑜𝑢𝑡
= 𝛼1 (𝑚̇ 𝑇𝑢𝑟𝑏 )2 + 𝛼2 (𝑚̇ 𝑇𝑢𝑟𝑏 ) + 𝛼3 21

Where, 𝛼3 is a constant with value 1. The rest of the 𝛼𝑖 terms are not constsant since

there is a dependency on turbine inlet temperature and the mass flow rate. Based on the

turbine map, the 𝛼𝑖 terms were identified to be linear and second order polynomial terms

of turbine inlet temperature as follows.

𝛼1 = 𝛽1 𝑇𝑇𝑖𝑛 + 𝛽2 22

2
𝛼2 = 𝛽3 𝑇𝑇𝑖𝑛 + 𝛽4 𝑇𝑇𝑖𝑛 + 𝛽5 23

Where, the 𝛽𝑖 terms are constants identified by minimizing the error between the model

and the manufacturer data. The fit of the model vs the manufacturer’s data is shown for

four different turbine inlet temperatures in Figure III-7.

33
Figure III-7. Turbine pressure ratio as a function of turbine mass flow rate with map

points for four different turbine inlet temperatures.

34
3.1.4 Turbine efficiency model

In order to derive turbine isentropic efficiency, the normalized parameter blade

speed ratio [65] is defined as follows.

𝜔𝑇 𝑟𝑡
𝐵𝑆𝑅 = 24
𝐾𝐸𝑥ℎ −1
𝑃 𝐾𝐸𝑥ℎ
√2𝐶𝑃 𝐸𝑥ℎ 𝑇𝑇𝑖𝑛 (1−( 𝑇𝑖𝑛 ) )
𝑃𝑇𝑜𝑢𝑡

Where, 𝑐𝑝𝐸𝑥ℎ and 𝐾𝐸𝑥ℎ are the constant pressure specific heat capacity and the

ratio of specific heat capacities of the exhaust gas, modeled as polynomial functions of

exhaust temperature as follows.

𝜂𝑇𝑢𝑟𝑏 = 𝑑1 𝐵𝑆𝑅 2 ⁡ + ⁡ 𝑑2 𝐵𝑆𝑅 + 𝑑3 25

The parameters 𝑑𝑖 are obtained similarly to the previous polynomial approximations in

the compressor and turbine models. The turbine efficiency is shown in Figure III-8 as a

function of BSR. The model is accurate within five percent for majority of the turbine

map points except for the six points at the lowest turbo speed.

35
Figure III-8. Modeled turbine efficiency curve vs map points

3.1.5 Turbocharger torque models

The compressor and turbine torque models are given by the following equations [42]
𝐾−1
𝑃𝐵𝑠𝑡 𝐾 1
𝜏𝐶𝑜𝑚𝑝 = 𝑚̇𝐶𝑜𝑚𝑝 . 𝑐𝑝𝐴 . 𝑇𝐶𝑖𝑛 [(𝑃 ) − 1] . [𝜂 ] 26
𝐶𝑖𝑛 𝐶𝑜𝑚𝑝 .𝜔𝑇

𝐾𝐸𝑥ℎ −1
𝑃𝑇𝑖𝑛 𝐾𝐸𝑥ℎ 𝜂𝑇𝑢𝑟𝑏
𝜏 𝑇𝑢𝑟𝑏 = 𝑚̇ 𝑇𝑢𝑟𝑏 . 𝑐𝑝𝐸𝑥ℎ . 𝑇𝐸𝑥ℎ [1 − (𝑃 ) ].[ ] 27
𝑇𝑜𝑢𝑡 𝜔𝑇

These equations are used in the turbocharger rotational dynamics equation as follows

36
𝑑𝜔𝑇 1
= 𝐽 (𝜏 𝑇𝑢𝑟𝑏 − 𝜏𝐶𝑜𝑚𝑝 − 𝜏𝐹𝑟𝑖𝑐𝑡 ) 28
𝑑𝑡 𝑇

Where, 𝜏𝐹𝑟𝑖𝑐𝑡 is the friction torque which is assumed to be zero, since this

parameter is associated with the mechanical efficiency term which is lumped into the

turbine efficiency as mentioned in [53]. 𝐽𝑇 is the moment of inertia of the rotor of the

turbocharger. This parameter can be measured based on techniques outlined in [43].

Alternatively it can be modeled if the exact geometry and material composition of the

turbocharger rotor assembly is available. These parameters can be used on the part file in

any commercial computer aided design software to estimate the moment of inertia of the

entire rotor assembly. For this research, this parameter has been derived experimentally

by comparing the turbocharger speed behavior of the 1D Engine simulation model to the

experimental data for a transient engine response. The moment of inertia utilized in the

1D Model is adjusted till the 1D Engine model mimics the turbo speed behavior of the

real engine. The final value was determined to be 1.7⁡ × 10−5 𝑘𝑔/𝑚2 which is within

the range of values for similar sized turbochargers. The turbocharger speed response for

the tuned 𝐽𝑇 can be seen in Figure III-9 and Figure III-10. The time constants of the 1D-

Engine simulation as well as the experimentally measured turbocharger speed appear

very similar to each other.

37
Figure III-9. Comparison of measured and GT-Power modeled transient turbocharger

speed response with tuned inertia for 1500 RPM

Figure III-10. Comparison of measured and GT-Power modeled transient turbocharger

speed response with tuned inertia for 2500 RPM

38
3.2 Data driven parameter models for engine cylinders

Modeling of the cylinder gas-exchange and combustion processes has been an

active research area since the advent of internal combustion engines. Parametric models

based on physics, empirical relations and measured data have been widely developed and

investigated. Since, engine modeling is not the focus of this research and the means to

obtain measurement data is easily available, a data driven approach has been adopted to

model these processes. The data was obtained by operating the engine at multiple

actuator combinations at 2000 and 3000 RPM. Each combination was allowed to settle

for some time before transitioning to the next combination. This time period is critical

because lesser time results in some of the slower dynamic behavior like exhaust

temperature to be transient even at the point of transition to the next combination. More

time per operating point is desirable but would lead to increased time required to obtain

the entire data set. A time period of five seconds was chosen per operating combination

which resulted in a total of forty hours of dynamometer test time. The actuator

combinations are shown in Table III-2. The actuator combinations are programmed into

an automated test cycle and deployed onto the ES910’s real-time controller.

In the control oriented model of the engine, there are four models that correspond

to the gas-exchange and in-cylinder combustion process of the engine. The gas-exchange

model is the volumetric efficiency (𝜂𝑉𝐸 ) model The remaining cylinder process models

are used to predict the Net Indicated Mean Effective Pressure (𝐼𝑀𝐸𝑃𝑛 ), Coefficient of

39
Variation of 𝐼𝑀𝐸𝑃𝑛 (𝐶𝑂𝑉𝐼𝑀𝐸𝑃𝑛 ), Turbine inlet gas temperature (𝑇𝐸𝑥ℎ ), and the square of

the Knock Intensity (𝐾𝐼 2 ).

Table III-2. Actuator combinations used for data collection for ANN model training

Variable Points Units

RPM 2000, 3000 RPM

Load (WG) 0, 50, 100 %

EGR 0, 5, 10, 15, 20 %

Spark 10, 5, 0, -5, -10 deg from base

P_int (throttle) 25, 31, 37, 44 %

ICL 40, 30, 20, 10, 0 (bTDC)

ECL 5, 15, 25, 35, 45 (aTDC)

The data acquisition is done conveniently using ETAS INCA’s Open Hardware

Integration (OHI) interface. Cyclic, crank-angle resolved cylinder pressure measurement

is required for the calculation of 𝐼𝑀𝐸𝑃𝑛 and 𝐾𝐼 2 . AVL Indicom’s calc-graf tool is

utilized to calculate 𝐾𝐼 2 in real time. 𝐼𝑀𝐸𝑃𝑛 is automatically calculated by AVL Indicom

per cycle. The INCA OHI interface is configured to communicate in real-time with AVL

Indicom to receive a configured list of parameters computed from the cylinder pressure.

This list of parameters is then available in the INCA workspace which also consists of the

control actuator signals sent to the engine. This test structure shown in Figure III-11

40
facilitates collection of data in one file per test using only one software (INCA) which

makes post processing of this data much simpler.

Figure III-11. Experimental setup to facilitate data acquisition for engine cylinder models

Once data collection is complete, it is post-processed to identify the steady state

portion of the five seconds per combination of input actuators. This is done because the

measurement of 𝜂𝑉𝐸 requires measurement of mass flow rate across the intake valves

which is not done directly on the engine. Instead, the mass air flow rate sensor which is

far upstream of the cylinders is utilized, which during transient operation is not

representative of the engine mass flow rate.

Similarly, 𝐶𝑂𝑉𝐼𝑀𝐸𝑃𝑛 calculation requires an ensemble of 𝐼𝑀𝐸𝑃𝑛 measurements at

a steady state. Typically, this calculation is done based on hundreds of engine cycles of

steady state data. For the un-boosted test conditions upto 600 consecutive steady-state

41
cycles could be identified from the data. However, for some of the high-load boosted

conditions, the 𝐼𝑀𝐸𝑃𝑛 would still be in a slow transient at the end of the five seconds due

to the slow dynamics (exhaust manifold temperature, intercooler outlet gas temperature)

of the engine. For such conditions, the last twenty cycles of the engine operation before

the actuator combination change were considered to be steady-state operating conditions.

The inputs to the ANN models are given in Table III-3

Table III-3 Inputs used for training ANNs

Input Unit

Spark deg bTDC

Intake manifold pressure Bar

Intake temperature K

Intake valve phasing deg aTDC

Exhaust valve phasing deg aTDC

RPM rpm

Partial pressure of EGR bar

42
3.2.1 𝜼𝑽𝑬 model

Neural network based methods have been investigated for engine air-flow

estimation by Nicolau et al in [84] and Wu et al in [85]. Neural networks training is

performed using Matlab’s neural network toolbox [86]. The accuracy for the model for

the collected data can be seen in Figure III-12. This network is modeled using a fifteen

neuron ANN with a single hidden layer. Majority of the test data points fall within the +/-

10% band. As seen in Figure III-12

Figure III-12. Modeled vs measured volumetric efficiency [%] for all data with +/- 10%

error bands represented by dashed red line

43
3.2.2 𝑰𝑴𝑬𝑷𝒏 model

Similar to the volumetric efficiency model the IMEPn is modeled using a fifteen

neuron, single hidden layer ANN. The model seems to lose accuracy at the lower load

points due to the high combustion variability noticed at these conditions resulting in

noisy 𝐼𝑀𝐸𝑃𝑛 data.

Figure III-13. Modeled vs measured IMEPn for all data with +/- 10% error bands

represented by dashed red line

44
3.2.3 𝑪𝑶𝑽𝑰𝑴𝑬𝑷𝒏 model

The 𝐶𝑂𝑉𝐼𝑀𝐸𝑃𝑛 ⁡is computed as follows for steady state ensembles of data.

𝑠𝑡𝑑𝑑𝑒𝑣(𝐼𝑀𝐸𝑃𝑛 )
𝐶𝑂𝑉𝐼𝑀𝐸𝑃𝑛 = III-29
𝑚𝑒𝑎𝑛(𝐼𝑀𝐸𝑃𝑛 )

Where, 𝑠𝑡𝑑𝑑𝑒𝑣(𝐼𝑀𝐸𝑃𝑛 ) is the standard deviation of the 𝐼𝑀𝐸𝑃𝑛 and 𝑚𝑒𝑎𝑛(𝐼𝑀𝐸𝑃𝑛 ) is

the average 𝐼𝑀𝐸𝑃𝑛 . The model is able to capture trends but lacks accuracy.

Figure III-14. Modeled vs measured 𝐶𝑂𝑉𝐼𝑀𝐸𝑃𝑛 for all data with +/- 10% error bands

represented by dashed red line

45
3.2.4 𝑲𝑰𝟐 model

The 𝐾𝐼 2 model has the worst accuracy of all the models. This is primarily due to

the lack of training data in the heavy knocking region. Since knocking is very detrimental

to the engine, collection of data at heavy knock was limited.

Figure III-15. Modeled vs measured 𝐾𝐼 2 for all data with +/- 10% error bands

represented by dashed red line

46
3.2. 𝑻𝑬𝒙𝒉 model

Similar to the volumetric efficiency and 𝐼𝑀𝐸𝑃𝑛 models, the steady state exhaust

temperature model is able to predict most of the points within the +/- 10% region. The

model was fitted to steady state exhaust temperature because the time constant associated

with the transient exhaust temperature is much longer than the turbocharger time

constant. Including these dynamics in the control model would require lengthening of the

horizon and consequently increasing the sampling time of the MPC to maintain real-time

execution feasibility, which would then require removing some of the pressure dynamics

to keep number of states in the model low.

Figure III-16. Modeled vs measured steady state exhaust temperature for all data +/- 10%

error bands represented by dashed red line. The accuracy of each of these neural

networks is summarized in Table III-4

47
Table III-4. Accuracy of trained ANNs

Parameter RMS Error

𝐼𝑀𝐸𝑃𝑛 0.45 Bar

𝐶𝑂𝑉𝐼𝑀𝐸𝑃𝑛 1.4%

𝐾𝐼 2 0.297 Bar2

𝜂𝑉𝐸 6.7%

𝑇𝐸𝑥ℎ 14 Deg K

48
3.3 Continuous time state space model of the system

Based on the models described in the previous sections, a continuous time state-

space model of the system is formed. The inputs and outputs are also normalized in

magnitude to have single digit values during operation. The units which are used for

normalization are shown in Table III-5

Table III-5 Units used for normalization of state space model

State/input/parameter Unit

Spark timing deg bTDC/10

Mass flow rate kg/min

Intake valve phasing deg aTDC/10

Exhaust valve phasing deg aTDC/10

Pressure Bar

IMEPn Bar

COV [Ratio]

Turbo speed kRad/sec

Temperature Deg K

Volume m^3

Based on the physical unit based normalization scheme above the state dynamics and the

outputs are represented as follows.

49
 ̇ = − 𝑉𝑑. 𝜂𝑉𝐸 .𝑅𝑃𝑀.𝑃𝐼𝑚 +
𝑃𝐼𝑚
𝑅.𝑇𝐼𝑚
6
𝑚̇𝐼𝑇𝑉 30
𝑉 .120
𝐼𝑚 𝑉 𝐼𝑚 .6×10

𝑅(𝑇𝐶𝑜𝑚𝑝𝑜𝑢𝑡 +𝑇𝐼𝐶𝑂𝑢𝑡 )
̇ =(
𝑃𝐵𝑠𝑡 ) (𝑚̇𝐶𝑜𝑚𝑝 − 𝑚̇𝐼𝑇𝑉 − 𝛽𝐵𝑜𝑣 (𝑃𝐵𝑠𝑡 − 𝑃𝐶𝑖𝑛 )) 31
2𝑉𝐵𝑠𝑡

𝑉 𝜂𝑉𝐸 .𝑅𝑃𝑀.𝑃𝐸𝑔𝑟 𝑅.𝑇𝐼𝑚

̇
𝑃𝐸𝑔𝑟 = − 𝑑. 𝑉 .120 +𝑉 6
𝑚̇𝐸𝑔𝑟 32
𝐼𝑚 𝐼𝑚 .6×10

𝐽−1
𝑇
𝜔̇ 𝑇 = 1000 (𝜏 𝑇𝑢𝑟𝑏 − 𝜏𝐶𝑜𝑚𝑝 ) 33

𝑇
𝑥 = [𝑃𝐼𝑚 ⁡𝑃𝐵𝑠𝑡 ⁡𝑃𝐸𝑔𝑟 ⁡𝜔 𝑇 ] , 𝑥 ∈ ℝ4×1 34

Where, suffixes 𝐼𝑚, 𝐵𝑠𝑡⁡and 𝐸𝑔𝑟 mean Intake manifold, post-compressor boost

manifold, and the EGR in the intake manifold. The input vector to the model is defined as

follows.
𝑇
𝑢 = [𝑚̇𝐼𝑇𝑉 ⁡𝛽𝐵𝑜𝑣 ⁡⁡𝑚̇𝑊𝐺 ⁡𝑚̇𝐸𝑔𝑟 ⁡𝑆𝑃𝐾⁡𝐼𝐶𝐿⁡𝐸𝐶𝐿] , 𝑢 ∈ ℝ7×1 35

Where, suffixes 𝐼𝑇𝑉, 𝐵𝑜𝑣⁡and 𝑊𝐺 denote Intake throttle valve, Blow-off valve, and

Turbine waste-gate respectively. The output that is to be tracked and the constraints are

given by the following equations.

𝑦 = 𝐼𝑀𝐸𝑃𝑛 , 𝑦 ∈ ℝ 36

𝑇
𝑧 = [(𝐶𝑂𝑉𝐼𝑀𝐸𝑃𝑛 )⁡(𝐾𝐼 2 )⁡(𝑃𝐼𝑚 − 𝑃𝐵𝑠𝑡 )⁡(𝑚̇𝑆𝐿 − 𝑚̇𝐶𝑜𝑚𝑝 )⁡(𝑚̇𝑊𝐺 − 𝑚̇𝐸𝑛𝑔 )] , 𝑧 ∈ ℝ5×1

37

50
Where, the constraint 𝑃𝐼𝑚 − 𝑃𝐵𝑠𝑡 < 0 and 𝑚̇𝑊𝐺 − 𝑚̇𝐸𝑛𝑔 < 0⁡ensures that the flow across

the intake throttle valve and the waste-gate is physically meaningful. The constraint

𝑚̇𝑆𝐿 − 𝑚̇𝐶𝑜𝑚𝑝 < 0 ensures that the compressor mass flow rate is always on the right hand

side of the surge line modeled as a linear function of 𝑃𝐵𝑠𝑡 as follows similar to [88].

𝑚̇𝑆𝐿 = 𝑆𝐺1 . 𝑃𝐵𝑠𝑡 + 𝑆𝐺2 38

Where, 𝑆𝐺1 and 𝑆𝐺2 are the coefficients of the surge-line of the compressor.

Open-loop validation of the state space model against experimentally measured engine
data has been shown in Figure III-17 for 3000RPM. The complete error metrics for the
0D engine model validated against four different engine transients at 2000 and 3000
RPM are shown in
Table III-6. The model is within +/- 10% accuracy for the states and the IMEPn output.

Table III-6 Error metrics of the 0D model for four different transient engine trajectories.

Parameter RMS Mean

IMEPn 1.07 4.9%

Pim 0.01 8.25%

Pbst 0.029 7.1%

Pegr 0.00008 7.6%

Wt 2.39 9.15%

51
Figure III-17 Open-loop system trajectory for 0D model vs experimentally collected data

52
 IV. NON-LINEARITY ANALYSIS OF CONTROL ORIENTED MODEL

4.1 Local linear approximation

Utilization of the non-linear state-space model in MPC requires local linear

approximation . Previous research has used approximated linear state-space models for

utilization in MPC. The underlying assumption of these approaches is that the behavior of

the non-linear engine system is linear-like for the entirety of the prediction horizon of the

MPC. In order to understand the effect of this assumption on the performance of MPC,

the following analysis has been performed.

 The deviation of the locally linearized model from the non-linear model is quantified

as the Linear Approximation Error (LAE)

 The LAE has been evaluated for a sinusoidal as well as a step change in inputs

The analysis has been done for six different steady state operating points which are

combinations of 2000 and 3000 RPM and IMEPn load levels of 5, 10 and 15 Bar.

4.2 Step input transient testing

The assumption that the linearized model is a sufficiently accurate representative

of the non-linear model 30 to 37 of the system is often made while designing MPC

algorithms. The IMEPn accuracy of the linearized model is of specific importance for

MPC designed for torque reference tracking. The LAE for the IMEPn output has been

quantified by first linearizing the model at a steady state operating point. Then the open

53
loop inputs which correspond to a deviation from the steady state IMEPn are applied to

both the models. The LAE for IMEPn is then examined as a function of time-step N from

the linearization point as shown in Figure IV-1.

Figure IV-1 Deviation between the non-linear system model and the linearized system

model as a function of time step N from the linearization. The deviation is induced by

perturbing 𝑚̇𝐼𝑇𝑉 corresponding to commanded 𝐼𝑀𝐸𝑃𝑛

The IMEPn LAE is quantified as the absolute error between the IMEPn from the non-

linear and the linearized model as follows

𝛿⁡𝐼𝑀𝐸𝑃𝑛 = |⁡𝐼𝑀𝐸𝑃𝑛𝑁𝐿 − 𝐼𝑀𝐸𝑃𝑛𝐿 | 39

54
The 𝛿⁡𝐼𝑀𝐸𝑃𝑛 is computed for different open loop step input 𝑚̇𝐼𝑇𝑉 for which the final

values correspond to the steady-state deviation from the linearization point IMEPn. This

is denoted in general as ∆⁡𝐼𝑀𝐸𝑃𝑛. Figure IV-2 shows the 𝛿⁡𝐼𝑀𝐸𝑃𝑛 as a function of the

time-step after linearization N and ∆⁡𝐼𝑀𝐸𝑃𝑛 for all six operating conditions. As expected

for ∆⁡𝐼𝑀𝐸𝑃𝑛 = 0, all of the conditions show 𝛿⁡𝐼𝑀𝐸𝑃𝑛 = 0 since there is no change in

input. Also, as expected, 𝛿⁡𝐼𝑀𝐸𝑃𝑛 = 0 for 𝑁 = 0 since it corresponds to the time at

which the model is linearized. As ∆⁡𝐼𝑀𝐸𝑃𝑛 is deviated from 0 and N is increased,

𝛿⁡𝐼𝑀𝐸𝑃𝑛 begins to grow larger. The highest deviation is seen at the 3000 RPM 15 Bar

case for ∆⁡𝐼𝑀𝐸𝑃𝑛 = −4. This implies that the non-linear 𝐼𝑀𝐸𝑃𝑛 output model varies

more with time at this operating condition compared to the others. Similarly, the least

deviation is seen at 3000 RPM 5 Bar case. Implementing MPC using such approximated

models would result in possibly reduced optimality and constraint violation. Based on the

analysis done on the six operating conditions shown in the errors induced by linear

approximation are significantly different as a function of operating conditions. Additional

linearization within the horizon causes 𝛿⁡𝐼𝑀𝐸𝑃𝑛 to naturally reduce due to improved

local accuracy of the linearized model at the additional linearization points. However,

since model linearization is a computationally expensive process, it is necessary to

identify the number of times the non-linear model needs to be linearized in order to

maintain a prescribed model accuracy. Table IV-1. shows the minimum number of

linearizations required to maintain different minimum accuracy levels of modeled

𝐼𝑀𝐸𝑃𝑛.

55
56
Figure IV-2 Model error as a function of commanded open-loop change in IMEPn

shows that the deviation between linearized and non-linear model varies as a function of

operating condition and the time –step after the point of linearization

Table IV-1 Minimum number of model linearizations required in horizon to maintain

IMEPn model accuracy for varying input frequency

2000 RPM 3000 RPM Max IMEPn error

[%]

15 Bar 1 2 5

10 Bar 1 1 5

5 Bar 10 2 5

15 Bar 1 5 3

10 Bar 2 2 3

5 Bar >10 5 3

15 Bar 5 5 1

10 Bar 5 5 1

5 Bar >10 10 1

57
4.3 Sinusoidal input transient testing

Similar to the step-input tests, sinusoidal inputs of frequency increasing from 0 to 2

Hz was applied to the model to understand the LAE as a function of input frequency. The

significance of this test is to analyze the dependence between the LAE and the rate at

which the engine transient occurs. Similar to the step input test, Table IV-2 shows the

𝛿⁡𝐼𝑀𝐸𝑃𝑛 as a function of the time-step after linearization N and frequency of the input

that was applied to the models. At every operating condition a certain frequency results in

the maximum LAE. This frequency ranges between 0.5 and 1.5 Hz depending on the

operating condition. However, similar to the step-input test the minimum 𝛿⁡𝐼𝑀𝐸𝑃𝑛 is

seen at the 3000 RPM 5 Bar case. The maximum 𝛿⁡𝐼𝑀𝐸𝑃𝑛 is seen at the 3000 RPM 10

Bar case for 0.9 Hz input frequency. Similar to the step-input case the minimum number

of linearizations required for different prescribed accuracy levels have been quantified. In

In both of the tests we can see that some of the operating conditions require the model to

be linearized more than 10 times within the horizon to meet the desired accuracy level.

This is due to the inaccuracy associated with discretization of the continuous model.

Since the sample time chosen here is 50ms, for a prediction horizon of 500ms, the

maximum number of linearizations that can be performed is 10.

58
Table IV-2 Minimum number of model linearizations required in horizon to maintain

IMEPn model accuracy for varying input frequency

2000 3000 Max IMEPn

RPM RPM error [%]

15 Bar 1 2 5

10 Bar 2 5 5

5 Bar >10 5 5

15 Bar 2 5 3

10 Bar 5 5 3

5 Bar >10 5 3

15 Bar 10 5 1

10 Bar 10 10 1

5 Bar >10 >10 1

59
Figure IV-3 Model error as a function of frequency of commanded open-loop model

input shows that the deviation between linearized and non-linear model generally peaks

between 1.5 Hz and 0.5 Hz depending on operating condition

60
 The choice of prediction model utilized for MPC is critical for the stability,

optimality and robustness of the closed loop controller performance. This research

presents a method of evaluating the impact of local linear approximation error on the

modeled IMEPn by manipulation of the throttle mass flow rate input. The same method

could be applied to every other input of the model to quantify the effect on the rest of the

outputs of the model.

The transient tests conducted in this study show clear deviation between the non-

linear model and its linearized version for different operating conditions. Based on this

analysis it is clear that this deviation varies significantly with the type of transient

induced in the system. The step transients are generally more forgiving with respect to

the number of model linearizations required, compared to the sinusoidal transients. The

sinusoidal transients clearly show a narrow frequency range where the deviation in the

models is highest.

In order to maintain a specific minimum accuracy, it is possible to identify the

minimum number of linearizations required within the horizon as a function of operating

condition and the commanded transient to the MPC. However, for some operating

conditions, linearization even at every stage within the prediction horizon is not sufficient

to ensure that Max IMEPn error stays within 5%

61
 V. NON-LINEAR MODEL PREDICTIVE CONTROL DESIGN

5.1 Horizon length and sample time determination

As mentioned in [54] by Zhu, a cascaded control structure is suitable for engine

control. Complex cycle-cycle behavior and fast dynamics can be handled effectively by

simpler, computationally cheaper controllers. However, the determination of the MPC

sample time 𝛿𝑡 is dependent on the engine dynamics considered in the MPC control

model, the prediction/control horizon considered and the computational ability of the

micro-controller on which the MPC is designed to execute on as shown in Figure V-1.

Closed loop stability of MPC requires prediction horizon to be long enough to

accommodate the slowest dynamics in the system [55][56]. The turbocharger speed has

the slowest dynamics in the system with time constant in the order of seconds (ignoring

the exhaust manifold wall temperature and the intercooler outlet temperature dynamics).

Figure V-1 Approximate relationship between horizon length considered for MPC,

size of the model and the computational expense of real-time execution.

62
Figure V-2 The selection of MPC sample time and prediction horizon time dictates the

choice of appropriate engine dynamics in the engine model

The sample time is chosen to be 50 ms in order to have a MPC prediction/control

horizon length of 10 steps which is 0.5 s. This allows consideration of the turbocharger

dynamics, the filling and emptying dynamics of the boost manifold and the intake

manifold. The exhaust manifold and the other control volumes on the engine are too

small and hence have almost instantaneous dynamics which can then be handled by lower

level controllers.

Based on the study shown in the previous chapter, it is evident that utilizing one

linearized model to predict the behavior of the engine over a period of time similar to the

prediction horizon of the MPC can result in significant errors in the modeled outputs.

Ideally, linearization at every stage within the prediction horizon would be desirable to

conserve accuracy. However, the process of obtaining the linearized model requires

63
computational resources. The following section will explore the methods used to

minimize computational burden associated with linearization.

5.2 Non-linear system model linearization

The system model described in (30) to (37) can be linearized as follows using the

Taylor’s series expansion (ignoring higher terms) in continuous time domain as follows.

𝜕𝑓 𝜕𝑓
𝑥̇ = 𝑓(𝑥0 , 𝑢0 ) + 𝛿𝑥̇ = 𝑥̇ 0 + 𝜕𝑥 | 𝛿𝑥 + 𝜕𝑢| 𝛿𝑢 40
𝑥0 ,𝑢0 𝑥0 ,𝑢0

𝜕𝑔 𝜕𝑔
| |
𝑦 𝑦0 𝜕𝑥 𝑥0 ,𝑢0 𝜕𝑢 𝑥0 ,𝑢0
[ ] = [ 𝑧 ] + [𝜕ℎ ] 𝛿𝑥 + [𝜕ℎ ] 𝛿𝑢 41
𝑧 0 | |
𝜕𝑥 𝑥0 ,𝑢0 𝜕𝑢 𝑥0 ,𝑢0

Where, 𝑥0 and 𝑢0 are the nominal state vectors and input vectors respectively. 𝑦0

and 𝑧0 are the corresponding nominal tracking output and constraint output vectors. 𝑥̇ 0 is

the vector of derivatives of the states evaluated at the nominal point. 𝛿𝑥 is the deviation

𝜕𝑓 𝜕𝑓
of the states from the nominal point. | , | are the Jacobian matrices evaluated
𝜕𝑥 𝑥0 ,𝑢0 𝜕𝑢 𝑥0 ,𝑢0

at 𝑥0 and 𝑢0 by partial differentiation of the state dynamics 3-30 to 3-33 with respect to 𝑥

𝜕𝑔 𝜕𝑔 𝜕ℎ 𝜕ℎ
and 𝑢. Similarly, 𝜕𝑥 | , 𝜕𝑢| , 𝜕𝑥 | and 𝜕𝑢| are the Jacobian matrices
𝑥0 ,𝑢0 𝑥0 ,𝑢0 𝑥0 ,𝑢0 𝑥0 ,𝑢0

evaluated at 𝑥0 and 𝑢0 by partial differentiation of the output functions 3-36 to 3-37 with

respect to 𝑥 and 𝑢.

64
5.2.1 Drawbacks of numerical linearization

These partial differential functions described above can be approximated by using

the forward finite difference approach as follows.

𝜕𝑐(𝑥) (𝑐(𝑥+Δ𝑥)−𝑐(𝑥))
≈ ⁡ 42
𝜕𝑥 Δ𝑥

Where, Δ𝑥 is the perturbation magnitude which is chosen to be sufficiently small

to approximate the slope of the arbitrary non-linear function 𝑐. Utilizing this method is

simple since it only requires evaluations of 𝑐 at different values of the argument.

However, since the engine system model has four states and seven inputs, utilizing this

method to obtain the linearized system model at one nominal operating condition will

require thirty eight evaluations of the system dynamics 3-30 to 3-33. Similarly, in order

to obtain the linearized output models, the output functions 3-36 to 3-37 would have to be

evaluated fifty six times. This would lead to ninety four evaluations of the system model

in order to obtain the linearized system dynamics. This process would have to be repeated

at every stage in the prediction horizon in order to minimize the linear approximation

error, resulting in nine hundred and forty evaluations of the system model at every time-

step of execution. This process takes a large amount of memory and processor throughput

and leaves no resources left for the optimization algorithm. In the worst case, it will cause

a task overrun wherein the computation fails to execute within the given time-step.

65
5.2.2 Hybrid linearization of the system model

The advantage of utilizing analytical differentiation is that unlike the numerical

approach described in the previous section, the analytical solution is exact and it requires

significantly less number of executions of the system model. A series of steps taken in

order to maximize analytical differentiation within the system model is described in the

following sections.

5.2.2.1 Fixed zero identification

The system model is first analyzed for potential to apply analytical partial

differentiation. In this the first process is to identify the fixed zeros in the system model.

𝜕𝑓𝑃𝐼𝑚 𝜕𝑓𝑃𝐼𝑚
𝟎 𝟎
𝜕𝑃𝐼𝑚 𝜕𝑃𝐸𝑔𝑟
𝜕𝑓𝑃𝐵𝑠𝑡 𝜕𝑓𝑃𝐵𝑠𝑡
𝟎 𝟎
𝜕𝑓 𝜕𝑃𝐵𝑠𝑡 𝜕𝜔𝑇
𝜕𝑥
= 𝜕𝑓𝑃𝐸𝑔𝑟 𝜕𝑓𝑃𝐸𝑔𝑟 43
𝟎 𝟎
𝜕𝑃𝐼𝑚 𝜕𝑃𝐸𝑔𝑟
𝜕𝑓𝜔𝑇 𝜕𝑓𝜔𝑇 𝜕𝑓𝜔𝑇 𝜕𝑓𝜔𝑇
[ 𝜕𝑃𝐼𝑚 𝜕𝑃𝐵𝑠𝑡 𝜕𝑃𝐸𝑔𝑟 𝜕𝜔𝑇 ]

66
 𝜕𝑓𝑃𝐼𝑚 𝜕𝑓𝑃𝐼𝑚 𝜕𝑓𝑃𝐼𝑚
𝜕𝑚̇𝐼𝑇𝑉
𝟎 𝟎 𝟎 𝟎 𝜕𝐼𝐶𝐿 𝜕𝐸𝐶𝐿
𝜕𝑓𝑃𝐵𝑠𝑡 𝜕𝑓𝑃𝐵𝑠𝑡
𝜕𝑚̇𝐼𝑇𝑉 𝜕𝛽𝐵𝑂𝑉
𝟎 𝟎 𝟎 𝟎 𝟎
𝜕𝑓
= 𝜕𝑓𝑃𝐸𝑔𝑟 𝜕𝑓𝑃𝐸𝑔𝑟 𝜕𝑓𝑃𝐸𝑔𝑟 44
𝜕𝑢
𝟎 𝟎 𝟎 𝜕𝑚̇𝐸𝑔𝑟
𝟎 𝜕𝐼𝐶𝐿 𝜕𝐸𝐶𝐿
𝜕𝑓𝜔𝑇 𝜕𝑓𝜔𝑇 𝜕𝑓𝜔𝑇 𝜕𝑓𝜔𝑇
[ 𝟎 𝟎 𝜕𝑚̇𝑊𝐺
𝟎 𝜕𝑆𝑃𝐾 𝜕𝐼𝐶𝐿 𝜕𝐸𝐶𝐿 ]

𝜕𝑔 𝜕(𝐼𝑀𝐸𝑃𝑛 ) 𝜕(𝐼𝑀𝐸𝑃𝑛 )
=[ 𝟎 𝟎] 45
𝜕𝑥 𝜕𝑃𝐼𝑚 𝜕𝑃𝐸𝑔𝑟

𝜕(𝐶𝑂𝑉𝐼𝑀𝐸𝑃𝑛 ) 𝜕(𝐶𝑂𝑉𝐼𝑀𝐸𝑃𝑛 )
𝜕𝑃𝐼𝑚
𝟎 𝟎
𝜕𝑃𝐸𝑔𝑟
𝜕(𝐾𝐼 2 ) 𝜕(𝐾𝐼 2 )
𝟎 𝟎
𝜕𝑃𝐼𝑚 𝜕𝑃𝐸𝑔𝑟
𝜕ℎ
= 1 −1 𝟎 𝟎 46
𝜕𝑢
𝜕(𝜉𝑆𝑢𝑟𝑔𝑒 ) 𝜕(𝜉𝑆𝑢𝑟𝑔𝑒 )
𝟎 𝟎
𝜕𝑃𝐵𝑠𝑡 𝜕𝜔𝑇
𝜕(𝜉𝑊𝐺 ) 𝜕(𝜉𝑊𝐺 )
𝟎 𝟎
[ 𝜕𝑃𝐼𝑚 𝜕𝑃𝐸𝑔𝑟 ]

The fixed zeros are identified by simply inspecting the differential equations in

state dynamics and the output equations. A fixed zero would simply denote the absence

of the corresponding state or output within that equation/model. It is evident that there is

a significant number of fixed zeros in the model which substantially alleviates the

computational burden of model linearization. Since the forward throttle flow constraint

𝜕ℎ
𝜉𝐼𝑇𝑉 is a linear function of the states the corresponding elements of 𝜕𝑢 are 1 and -1

respectively.

67
5.2.2.2 Analytical and hybrid differentiation of system model

The analytical differentiation of the system dynamics for 𝑃𝐼𝑚 and 𝑃𝐸𝐺𝑅 is simple

and given by the following equations.

𝜕𝑓𝑃𝐼𝑚 𝑉 𝑅𝑃𝑀 𝜕𝜂
𝜕𝑃𝐼𝑚
𝑑
= −120𝑉 (𝑃𝐼𝑚 𝜕𝑃𝑉𝐸 + 𝜂𝑉𝐸 ) 47
𝐼𝑚 𝐼𝑚

𝜕𝑓𝑃𝐼𝑚 𝑉 𝑅𝑃𝑀 𝜕𝜂
𝑑
= −120𝑉 (𝑃𝑖𝑚 𝜕𝑃 𝑉𝐸 ) 48
𝜕𝑃𝐸𝑔𝑟 𝐼𝑚 𝐸𝑔𝑟

𝜕𝑓𝑃𝐼𝑚 𝑇𝐼𝑚 𝑅
=𝑉 6
49
𝜕𝑚̇𝐼𝑇𝑉 𝐼𝑚 6×10

𝜕𝑓𝑃𝐼𝑚 𝑉𝑑 𝑅𝑃𝑀𝑃𝐼𝑚 𝜕𝜂𝑉𝐸

= ( 𝜕𝐼𝐶𝐿 ) 50
𝜕𝐼𝐶𝐿 −120𝑉𝐼𝑚

𝜕𝑓𝑃𝐼𝑚 𝑉𝑑 𝑅𝑃𝑀𝑃𝐼𝑚 𝜕𝜂𝑉𝐸

= (𝜕𝐸𝐶𝐿 ) 51
𝜕𝐸𝐶𝐿 −120𝑉𝐼𝑚

𝜕𝑓𝑃𝐸𝑔𝑟 𝑉 𝑅𝑃𝑀 𝜕𝜂
𝑑
= −120𝑉 (𝑃𝐸𝑔𝑟 𝜕𝑃𝑉𝐸 ) 52
𝜕𝑃𝐼𝑚 𝐼𝑚 𝐼𝑚

𝜕𝑓𝑃𝐸𝑔𝑟 𝑉 𝑅𝑃𝑀 𝜕𝜂
𝑑
= −120𝑉 (𝑃𝐸𝑔𝑟 𝜕𝑃 𝑉𝐸 + 𝜂𝑉𝐸 ) 53
𝜕𝑃𝐸𝑔𝑟 𝐼𝑚 𝐸𝑔𝑟

𝜕𝑓𝑃𝐸𝑔𝑟 𝜕𝑓𝑃
= 𝜕𝑚̇ 𝐼𝑚 54
𝜕𝑚̇𝐸𝑔𝑟 𝐼𝑇𝑉

𝜕𝑓𝑃𝐸𝑔𝑟 𝑉𝑑 𝑅𝑃𝑀𝑃𝐸𝐺𝑅 𝜕𝜂𝑉𝐸

= ( 𝜕𝐼𝐶𝐿 ) 55
𝜕𝐼𝐶𝐿 −120𝑉𝐼𝑚

𝜕𝑓𝑃𝐸𝑔𝑟 𝑉𝑑 𝑅𝑃𝑀𝑃𝐸𝐺𝑅 𝜕𝜂𝑉𝐸

= (𝜕𝐸𝐶𝐿 ) 56
𝜕𝐸𝐶𝐿 −120𝑉𝐼𝑚

68
 It is visible in this analytical framework that there are common computations

𝜕𝜂𝑉𝐸
shared between the models. E.g. 𝜕𝐼𝐶𝐿
is common in (50) and (55) and is computed only

once, further reducing the computational load. For 𝑃𝐵𝑠𝑡 a hybrid approach is utilized. The

𝜕𝑓𝑃𝐵𝑠𝑡 𝜕𝑓𝑃𝐵𝑠𝑡
numerical differentiation as shown in (42) is used for and ⁡because of the
𝜕𝑃𝐵𝑠𝑡 𝜕𝜔𝑇

saturation functions and singularities in the 𝜓⁡functions of the compressor model.

However, the linearization w.r.t. the inputs is still done analytically as follows.

𝜕𝑓𝑃𝐵𝑠𝑡 𝑇𝐵𝑠𝑡 𝑅
= −𝑉 6
57
𝜕𝑚̇𝐼𝑇𝑉 𝐵𝑠𝑡 6×10

𝜕𝑓𝑃𝐵𝑠𝑡 𝑇𝐵𝑠𝑡 𝑅
= (𝑃𝐵𝑠𝑡 − 𝑃𝐶𝑖𝑛 ) (⁡−𝑉 6
) 58
𝜕𝛽𝐵𝑂𝑉 𝐵𝑠𝑡 6×10

The analytical differentiation of the 𝜔 𝑇 dynamics results in long, complicated

solutions which are marginally faster than the numerical solutions obtained from. Hence,

the numerical approach was utilized in obtaining these partial derivatives but after some

simplification as follows.

𝜕𝑓𝜔𝑇 1 𝜕𝜏
= 1000𝐽 ( 𝜕𝑃𝑇𝑢𝑟𝑏) 59
𝜕𝑃𝐼𝑚 𝑇 𝐼𝑚

𝜕𝑓𝜔𝑇 −1 𝜕𝜏𝐶𝑜𝑚𝑝
= 1000𝐽 ( ) 60
𝜕𝑃𝐵𝑠𝑡 𝑇 𝜕𝑃𝐵𝑠𝑡

69
 𝜕𝜏𝐶𝑜𝑚𝑝 𝜕𝜏𝑇𝑢𝑟𝑏
Where, the numerical computation associated with and is eliminated
𝜕𝑃𝐼𝑚 𝜕𝑃𝐵𝑠𝑡

due to the absence of the corresponding states in these equations, similar to the fixed

zeros.

5.2.2.3 Analytical differentiation of artificial neural network sub-models

Majority of the functions used to generate the tracking and constraint output are

the ANN models described in the previous chapter. Hence, the linearization of the

outputs requires linearization of the ANNs w.r.t. their inputs. A typical ANN model has

a graphical structure shown in Figure V-3.

Figure V-3. Graphical layout of Artificial Neural Network

70
This model can be mathematically represented as follows.

𝑦 = 𝑠(𝑊 𝑛 𝑠(𝑊 𝑛−1 𝑠(… 𝑠(𝑊 1 𝑥 + 𝑏1 ) + 𝑏 𝑛−1 ) + 𝑏 𝑛 ) 61

Where, 𝑊 𝑖 and 𝑏 𝑖 are the weight and bias vectors for the ith layer of the ANN

model respectively. 𝑥 is the input vector and n is the total number of layers in the

function. s is the sigmoid activation function of the following form. Hence as shown in

Figure V-3 the 𝛼 𝑖 terms are computed as follows.

𝛼1 = 𝑊 1 𝑥 + 𝑏1 V-62

𝛼 2 = 𝑊 2 𝑠(𝛼 1 ) + 𝑏2 V-63

𝛼 𝑛 = 𝑊 𝑛−1 𝑠(𝛼 𝑛−1 ) + 𝑏 𝑛−1 V-64

The function 𝑠 is the sigmoid activation function of the following form.

1
𝑠(𝛼) = 65
1+𝑒 −𝛼

Similar to the ‘Back Propagation’ algorithm used to train the ANN models, the

differential of this function w.r.t. its argument is computed as follows.

𝑠′(𝛼) = 𝑠(𝛼)(1 − 𝑠(𝛼)) 66

71
It is noteworthy that 𝑠(𝛼) is in fact computed during the forward execution of the ANN

model and is already available. Hence the partial derivative of the model with respect to

it’s inputs is easily computed using the chain rule as follows.

𝜕𝑦
= 𝑠 ′ (𝑊 𝑛 𝑠(𝑊 𝑛−1 𝑠(… 𝑠(𝑊 1 𝑥 + 𝑏1 ) + 𝑏 𝑛−1 ) + 𝑏 𝑛 )(𝑊 𝑛 )𝑇 𝑠 ′ (𝑊 𝑛−1 𝑠(… 𝑠(𝑊 1 𝑥 +
𝜕𝑥

𝑏1 ) + 𝑏 𝑛−1 ))(𝑊 𝑛−1 )𝑇 … 𝑠 ′ (𝑊 1 𝑥 + 𝑏1 )(𝑊 1 )𝑇 67

Where, (66) is utilized to compute 𝑠′ using results of the computation of 𝑠 from the

forward execution of the model. Figure V-4 shows the difference in execution time

between the analytical differentiation method shown above and the numerical

differentiation for different combinations of inputs. The execution of both methods was

performed on a desktop computer. The analytical method is over ten times faster than the

numerical approach.

72
 Figure V-4. Comparison between execution times of 𝐼𝑀𝐸𝑃𝑛 ANN model

linearization using numerical and analytical approach

Figure V-4 shows the difference in execution time between the numerical linearization

and the hybrid model linearization for the entire system model using the methods

described in this section. Similar to the difference in execution time for the ANN models

shown in Figure V-5. The execution time for the system model linearization using the

hybrid approach is also over ten times faster than the numerical approach.

73
 Figure V-5. Comparison between execution times of complete system model

linearization using numerical and analytical approach

5.3 Model discretization in time

Implementation of the system model in MPC requires discretization of the system

model in time. The non-linear model is open-loop stable in continuous time domain,

hence it is necessary to ensure that the discrete-time, linear model used for MPC is open-

loop stable as well to preserve accuracy. The manifold filling and emptying dynamics

have time constants which vary with engine load and speed. At high engine flow rates,

these time constants becomes significantly shorter than 𝛿𝑡. Such fast dynamics could be

74
eliminated from the model at these conditions, but that would require a variable-order

model which is not desirable for MPC formulation. Euler’s finite difference

approximation is a commonly used method for conversion of a model from continuous

time to discrete time domain. However using this method for the linear system leads to

numerical instability at some high load/speed conditions. . Utilizing the State Transition

Matrix (STM) method for time domain discretization generally preserves numerical

stability to a higher degree compared to the finite difference approach [83]. The discrete

time model is obtained using the following equation.

𝛿𝑡
𝛿𝑥𝑘+1 = 𝑒 𝐴𝛿𝑡 𝛿𝑥𝑘 + (∫0 𝑒 𝐴(𝛿𝑡−𝜏) 𝑑𝜏𝐵) 𝛿𝑢𝑘 68

Where, 𝑒 𝐴𝛿𝑡 and 𝑒 𝐴(𝛿𝑡−𝜏) ) are matrix exponentials computed using the Pade

approximation. Hence, the discrete time model of the following form is derived from the

STM method.

𝛿𝑥𝑘+1 = 𝐴𝑑 𝛿𝑥𝑘 + 𝐵𝑑 𝛿𝑢𝑘 69

𝑦𝑘 𝑦0
[ 𝑧 ] = [ 𝑧 ] + 𝐶𝑑 𝛿𝑥𝑘 + 𝐷𝑑 𝛿𝑢𝑘 70
𝑘 0

Due to the nature of discrete-time process execution on an electronic micro-

controller, the commands generated by the MPC algorithm are available only at the

beginning of the next time step. This results in a unit time-step delay in the control

actions. However, the prediction model () has a non-zero D matrix which does not

75
account for this delay and also isn’t favorable for MPC-QP formulation. In order to

address this issue the discrete time model is modified as follows.

𝛿𝑥𝑘+1 𝐴 𝐵𝑑 𝛿𝑥𝑘 0
[ ]=[ 𝑑 ][ ] + [ ] 𝛿𝑢𝑘 71
𝛿𝑥̃𝑘+1 0 0 𝛿𝑥̃𝑘 𝐼

𝑦𝑘 𝑦0 𝛿𝑥𝑘
[ 𝑧 ] = [ 𝑧 ] + [𝐶𝑑 𝐷𝑑 ] [ ] 72
𝑘 0 𝛿𝑥̃𝑘

Augmentation of the original state vector with 𝑥̃𝑘 automatically induces a unit

step delay in the model and eliminates the 𝐷 matrix. The prediction model (71) to (72) is

time-varying. Hence, it is derived at every stage within the prediction horizon using the

method shown above and the set of these derived models is utilized to formulate the

MPC-QP.

5.4 MPC Cost function formulation

The formulation of the MPC-QP largely depends on the chosen objective function

that is to be minimized. A tracking objective function is chosen to formulate the MPC-QP

as follows.

𝑘+𝑁−1

𝐽(𝑥(𝑘), 𝑈(𝑘)) = ⁡𝑦̅(𝑘 + 𝑁) 𝑄𝑁 𝑦̅(𝑘 + 𝑁) + ∑ ⁡[𝑦̅(𝑖)𝑇 𝑄𝑦̅(𝑖) + 𝑢̅(𝑖)𝑇 𝑅𝑢̅(𝑖)]

𝑇

𝑖=𝑘

𝑦̅(𝑘) = ⁡𝑦(𝑘) − 𝑦𝑅𝑒𝑓 (𝑘)

𝑢̅(𝑘) = ⁡𝑢(𝑘) − 𝑢𝑅𝑒𝑓 (𝑘) 73

𝑈(𝑘) = [𝑢(𝑘), 𝑢(𝑘 + 1), … 𝑢(𝑘 + 𝑁 − 1)]𝑇

76
Where,

𝑄 ∈ ℝ>0 , 𝑄𝑁 ∈ ℝ>0 , 𝑅 ∈ ℝ7×7

>0

𝑄 and 𝑄𝑁 are the tuning matrices for the stage and terminal penalties on 𝐼𝑀𝐸𝑃𝑛

tracking error. 𝑄𝑁 is chosen large enough for closed loop stability [55][56]. 𝑅 is the

tuning matrix to penalize tracking error of the inputs and is always positive definite.

5.5 Reference input generation

The reference input 𝑢𝑅𝑒𝑓 is the input vector which minimizes fuel consumption

of the engine at steady state conditions. It is also utilized to generate the reference

trajectory of the prediction model around which the MPC finds the optimal solution.

𝑚̇𝐼𝑇𝑉𝑟𝑒𝑓 , 𝑚̇𝑊𝐺𝑟𝑒𝑓 and 𝑚̇𝐸𝐺𝑅𝑟𝑒𝑓 which are part of 𝑢𝑅𝑒𝑓 are particularly important to derive

the reference trajectory of the model in the prediction horizon. The steady state values of

𝑚̇𝐼𝑇𝑉𝑟𝑒𝑓 under boosted conditions can only be achieved by increasing the turbine power

by sufficiently reducing 𝑚̇𝑊𝐺 . Hence using the steady state 𝑚̇𝐼𝑇𝑉𝑟𝑒𝑓 can lead to a

reference trajectory which is physically impossible and undesirable. E.g. 𝑃𝐵𝑠𝑡 may drop

well below 𝑃𝐼𝑚 in the reference trajectory, even low enough to drop the compressor

pressure ratio below 1, resulting in choked operation. At this condition, the numerically

𝜕(𝑚̇𝐶𝑜𝑚𝑝 )
derived becomes 0, decoupling 𝑃𝐵𝑠𝑡 from the compressor mass flow rate model.
𝜕𝑃𝐵𝑠𝑡

This results in only 𝜔 𝑇 influence on the compressor mass flow rate. Since 𝜔 𝑇 has the

slowest dynamics, the compressor mass flow rate practically is ‘locked’ at the choked

position corresponding to 𝜔 𝑇 at that instant in the reference trajectory. For large tip-in

77
maneuvers, this problem results in an infeasible MPC-QP problem due to the imposed

𝜉𝐼𝑇𝑉 constraint. Similarly, utilizing the steady state 𝑚̇𝑊𝐺𝑟𝑒𝑓 and 𝑚̇𝐸𝐺𝑅𝑟𝑒𝑓 to obtain the

reference trajectory results in the 𝜉𝑊𝐺 and 𝐶𝑂𝑉𝐼𝑀𝐸𝑃𝑛 constraints being violated in the

reference trajectory. To address these problems in deriving 𝑢𝑅𝑒𝑓 , the following steps have

been taken. 𝑚̇𝐼𝑇𝑉𝑟𝑒𝑓 is derived using the orifice flow model and the corresponding steady

state throttle position. 𝑚̇𝑊𝐺𝑟𝑒𝑓 and 𝑚̇𝐸𝐺𝑅𝑟𝑒𝑓 are obtained as follows

𝑚̇𝑊𝐺𝑟𝑒𝑓 = 𝑊𝐺𝑟𝑎𝑡𝑖𝑜 (𝑚̇𝐼𝑇𝑉𝑟𝑒𝑓 ) 74

𝑚̇𝐸𝐺𝑅𝑟𝑒𝑓 = 𝐸𝐺𝑅𝑟𝑎𝑡𝑖𝑜 (𝑚̇𝐼𝑇𝑉𝑟𝑒𝑓 ) 75

5.6 MPC Quadratic program problem formulation and optimization

The non-linear constrained minimization problem is defined as follows.

min
𝑎𝑟𝑔⁡ 𝐽(𝑥(𝑘), 𝑈(𝑘))
𝑈(𝑘)

Subject to:

𝑧(𝑖) ≤ [0.15⁡1⁡0⁡0⁡0]𝑇

Based on the cost function Error! Reference source not found. and the prediction

model, an MPC - Quadratic Program is formulated as follows.

min 1 𝑇
𝑈 𝐻⁡𝑈 + 𝐿𝑇 𝑈⁡
𝑈 2

𝑀𝑈 ≤ 𝐹

Where,

𝐻 ∈ ℝ70×70
>0 , 𝐿 ∈ ℝ70 , 𝑀 ∈ ℝ70×70
>0 , 𝐹 ∈ ℝ70 , 𝑈 ∈ ℝ70

78
Solving the MPC-QP described above results in the optimal solution 𝑈 ∗ which is a

sequence of sub-optimal inputs in the vicinity of the reference inputs 𝑈𝑟𝑒𝑓 out of which

the input vector which corresponds to the current time step 𝑢∗ (𝑘) is chosen and applied

to the engine system.

5.7 Simulation with virtual engine in the loop

A 1D-dynamical engine model designed in GT-Suite was coupled with the control

system in order to perform co-simulation as shown in Figure V-6. Co-simulation setup

between GT-Power and SimulinkThe simulation was performed at 3000 RPM.

Figure V-6. Co-simulation setup between GT-Power and Simulink

79
Figure V-7 IMEPn response of MPC co-simulated with GT-Power

5.7.1 IMEPn tracking response

The closed-loop IMEPn response of the engine is shown in Figure V-7. It is

immediately visible that the MPC tracks the commanded IMEPn with some steady state

offset. This occurs due to the 𝑢𝑟𝑒𝑓 calibration being derived from the experimental engine

data. This reference input corresponds to a steady state IMEPn which is higher than the

desired value. Due to the tuning penalty⁡𝑅 on the reference inputs, the MPC generates

80
inputs which are a ‘trade-off’ between the steady-state reference inputs⁡𝑢𝑟𝑒𝑓 and the

inputs required to improve the tracking of the desired IMEPn.

Figure V-8 Control actuator trajectories for MPC cosimulation with GT-Power, air-path

actuators are on the left hand side and cylinder actuators are on the right hand side. Black

dashed lines are the steady-state reference inputs and blue lines are the inputs generated

by MPC

81
5.7.2 Air-path actuator trajectories

The control actuator trajectories for the IMEPn transient are shown in Figure V-8

The MPC shows differences in air-path actuation for the two tip-in maneuvers shown.

For the first tip-in at 2 seconds, the MPC commands minimal change in 𝑚̇𝐸𝑔𝑟 whereas,

for the second tip-in at 12 seconds the MPC commands a complete stoppage of 𝑚̇𝐸𝑔𝑟 for

a short period of time. 𝑚̇𝑊𝐺 is also stopped temporarily at the time of tip-in followed by a

return to it’s reference value. 𝛽𝐵𝑂𝑉 actuation on the second tip-in mitigates the drop in

𝑃𝐵𝑠𝑡 caused by the rapid increase of 𝑚̇𝐼𝑇𝑉 . For the tip-out at 7 seconds MPC mainly

commands 𝛽𝐵𝑂𝑉 to address the surge constraint.

5.7.3 Engine-cylinder actuators

The 𝑆𝑃𝐾 (Spark timing) actuator is significantly retarded from the reference

value during both the tip-in maneuvers. MPC also advances the intake cam position 𝐼𝑉𝑂

during the tip-in maneuvers for volumetric efficiency and air-flow improvement. During

the tip-out the MPC advances 𝑆𝑃𝐾 to handle the 𝐶𝑂𝑉𝐼𝑀𝐸𝑃𝑛 constraint.

5.7.4 Constraint handling

The last three constraint outputs in the constraint vector are purely air-path

constraints. Out of these constraints the constraints 𝑃𝐼𝑚 − 𝑃𝐵𝑠𝑡 < 0 and

𝑚̇𝑊𝐺 − 𝑚̇𝐸𝑛𝑔 < 0 are always satisfied on the GT-Power model independent of the MPC

commands. These constraints are in place merely to ensure physically feasible behavior

in the control oriented model (30) to (37). The surge constraint handling is demonstrated

82
in Figure V-9 where, the compressor operation momentarily crosses the surge-line but

returns back to the normal operating region of the compressor.

Figure V-9 Compressor operation during co-simulation of GT-Power with MPC

Since the knock model is not tuned and the 𝐶𝑂𝑉𝐼𝑀𝐸𝑃𝑛 cannot be evaluated using the GT-

Power model, they will be evaluated in the experimental testing section.

83
 VI. LOWER-LEVEL CONTROL AND ESTIMATION

The commands generated by MPC require conversion into actuator set-points for

real-time application. This lower-level actuation is categorized into two sections, air-path

control and engine-cylinders control. Real time estimation of the values of the states is

also required in order to formulate the MPC-QP problem.

6.1 Valve mass flow rate control

The air-path control consists of resolving the valve mass flow rate commands 𝑚̇𝐼𝑇𝑉 , 𝛽,

𝑚̇𝑊𝐺 , and ⁡𝑚̇𝐸𝐺𝑅 into valve position set-points using the orifice flow equation as follows.

𝑚̇√𝑅𝑇
𝐴𝑒 = 𝑃𝐼𝑛 76
𝜓( )𝑃
𝑃𝑂𝑢𝑡 ⁡ 𝐼𝑛

𝜃 = 𝑙𝑢𝑡(𝐴𝑒) 77

Where, 𝑙𝑢𝑡 is the lookup table function to convert commanded valve area 𝐴𝑒 to a valve

position 𝜃. 𝜓 is the flow function which has an infinite gradient when the pressure

difference ∆𝑃 across the valve approaches as shown in Figure VI-1. This results in

‘chattering’ in the valve position due to small changes in ∆𝑃.

84
Figure VI-1. The gradient of the flow function becomes very steep as the pressure

difference across the valve diminishes

In order to address this issue a simple filter with the following structure is applied.

𝜃𝐹𝑖𝑛𝑎𝑙 (𝑘) = 𝑤(𝑘)𝜃𝐼𝑛𝑖𝑡 (𝑘) + (1 − 𝑤(𝑘)) 78

𝑤(𝑘) = |𝑔∆𝑃(𝑘)| 79

Where, 𝑔 is a tuning parameter which can be tuned heuristically to mitigate chattering at

∆𝑃(𝑘) ≈ 0 while also having sufficiently fast transient response. The engine-cylinders

control is simply the realization of the commands 𝑆𝑃𝐾, 𝐼𝑉𝑂, and⁡𝐸𝑉𝐶

85
6.2 Low-pressure EGR ∆𝑷⁡control

Recent findings show the difficulty in control of mass flow rate across the LPE

system [37][38]. The standard orifice model accuracy deteriorates rapidly in a highly

pulsating ∆𝑃𝐿𝑃𝐸 ≈ 0. Increasing ∆𝑃𝐿𝑃𝐸 to atleast 5 kPa was observed to maintain the

orifice flow model’s accuracy within 5%. Hence, the 𝑃𝐶𝑇 valve is actuated using𝐴𝑒=

𝑚̇√𝑅𝑇
𝑃𝐼𝑛 (76). The mass air-flow sensor unit which is shortly upstream of the 𝑃𝐶𝑇
𝜓( )𝑃
𝑃𝑂𝑢𝑡 ⁡ 𝐼𝑛

valve is used for the 𝑚̇ and 𝑃𝐼𝑛 input. 𝑃𝑂𝑢𝑡 is simply set to 𝑃𝐼𝑛 − 5⁡𝑘𝑃𝑎. Since, the

turbine outlet pressure is always higher than 𝑃𝐼𝑛 , ∆𝑃𝐿𝑃𝐸 ≥ 5⁡𝑘𝑃𝑎 is guaranteed. However,

this method leads to much higher ∆𝑃𝐿𝑃𝐸 at higher power conditions where the ‘natural’

∆𝑃𝐿𝑃𝐸 without PCT throttling is close to 5 kPa. This method results in unnecessary

compressor inlet throttling, potentially limiting peak torque of the engine. Although this

is a caveat in this method, it is simple to implement and ensures acceptable accuracy in

controlling the mass flow rate through the LPE system.

6.3 Turbo speed estimation and state measurement

Since the prediction model is physics based, the states have physical meaning as

well. 𝑃𝐼𝑚 and 𝑃𝐵𝑠𝑡 measurements are part of the production engine sensor set. An intake

oxygen sensor is installed at the inlet of the intake manifold for real time measurement of

𝑃𝐸𝑔𝑟 . These measurements are also low pass filtered to minimize noise and oscillations

induced from the discrete-event gas exchange processes. The cutoff frequency high

86
enough to maintain transient accuracy within 50ms. 𝜔 𝑇 is estimated using the Fixed Point

Iteration method described in [89]. The compressor model is utilized for estimation.

̂𝐶𝑜𝑚𝑝 )
𝛼(Π𝑐𝑜𝑚𝑝 −Π ̂̇ 𝐶𝑜𝑚𝑝 )
(1−𝛼)(𝑚̇𝐶𝑜𝑚𝑝 −𝑚
̂ 𝑇 (𝑘) = 𝜔
𝜔 ̂ 𝑇 (𝑘 − 1) + 𝐾 ( ̂
𝜕Π
+ ̂̇
𝜕𝑚
) 80
𝐶𝑜𝑚𝑝 𝐶𝑜𝑚𝑝
̂𝑇
𝜕𝜔 ̂𝑇
𝜕𝜔

𝑃𝐵𝑠𝑡
Π𝐶𝑜𝑚𝑝 = 81
𝑃𝐶𝑖𝑛

1
𝛼= ̂ 𝐶𝑜𝑚𝑝
𝜕Π
82
1− ̂
𝜕𝑚̇𝐶𝑜𝑚𝑝

̂ 𝐶𝑜𝑚𝑝 and 𝑚
Where, Π ̂̇ 𝐶𝑜𝑚𝑝 is the predicted pressure ratio and compressor mass flow rate

̂ 𝑇 (𝑘 − 1) from the previous iteration and the

using the estimated turbo speed 𝜔

compressor model. 𝛼 dictates how much emphasis is placed on utilizing Π𝐶𝑜𝑚𝑝 over

̂ 𝑇 (𝑘). Figure VI-2 shows the estimated vs measured

𝑚̇𝐶𝑜𝑚𝑝 for estimation of 𝜔

turbocharger speed for a fully transient engine test. Some of the scatter in the Figure VI-2

is caused by the noise in measurement of turbocharger speed. The accuracy of the turbo

speed estimation using this method was observed to be within the 15% error band even

for low speed extrapolated region of the compressor map. The RMS error and max error

for this test is 0.28 kRad/s and 1.5 kRad/s respectively.

87
Figure VI-2.Real-time estimated vs measured turbocharger speed using fixed point

iteration method

88
 VII. EXPERIMENTAL VALIDATION OF MPC

7.1 Experimental control system hardware configuration

The architecture of the Rapid-control prototyping system interfaced with the

engine and it’s ECU is shown in Figure VII-1. The MPC derived in the previous section

requires higher computational ability compared to standard engine controllers. In order to

maintain real-time execution ability using this formulation, a Dspace DS1006 has been

utilized to execute the MPC program similar to [90][91]. The DS1006 executes in two

timer-task mode. The MPC program operates at the sample time 𝛿𝑡. At the end of each

step of it’s execution 𝑢∗ is sent to a communication block which executes at a sample

time of 2ms. This block consists of the CAN transmit subsystems which send 𝑢∗ to the

ES910 over a 1Mbit/s J1939 CAN bus. The same channel is used to receive states 𝑥 and

parameters 𝑘 from the ES910. These are utilized to execute the next step of the MPC

program. During testing, the message transmission and receive time was observed to be

less than 2ms which is negligible compared to the unit-step delay induced from discrete

time execution.

The lower-level control program is deployed onto an ETAS ES910 real-time

simulation controller. Because of the additional actuators and sensors associated with LP-

EGR and WG control an ES930 I/O extension unit has also been added to the ES910

controller. The 𝑃𝐸𝑔𝑟 measurement is also performed using an Oxygen sensor controller

box which outputs analog voltage as a function of Oxygen concentration. The lower

level control layer also has the position controllers for the LP-EGR, PCT and BOV

89
valves. These are conventional PID based controllers. Because of the faster dynamics of

the lower-level actuators, the sample time of the low- level controller and estimator is set

to 1ms.

Figure VII-1. Rapid control prototyping setup for experimental validation of MPC

90
7.2.1 Experimental test results 2900 RPM: Aggressive tip-in and tip-out maneuvers

Analysis of the MPC controller performance is done by comparing against the

open-loop fuel-economical control input 𝑢𝑅𝑒𝑓 . The analysis is done for step commands of

𝐼𝑀𝐸𝑃𝑛 at 3000 RPM to show the tip-in and tip-out response of both the control

approaches. Figure VII-2 shows the tip-in response of the controllers. It is evident that the

tip-in response of the MPC is much quicker than the open-loop control. The MPC

response also has a slight overshoot followed by stable steady state tracking. The air-path

actuator trajectories have been shown in Figure VII-3, Figure VII-4 and Figure VII-5.

Figure VII-2. Tip-in IMEPn response for MPC and Open-loop control strategy

At the instant of the tip-in the commanded reference 𝑚̇𝐼𝑇𝑉 is momentarily increased and

dropped based on the reference steady state intake throttle position which corresponded

to a pressure difference of approximately 15kPa across the intake throttle. Because of this

91
the MPC commanded 𝑚̇𝐼𝑇𝑉 impulse is slightly higher as the forward throttle flow

constraint is inactive in the reference trajectory. MPC completely stops 𝑚̇𝑊𝐺 and 𝑚̇𝐸𝐺𝑅

to accelerate turbocharger spool-up. 𝑚̇𝐸𝐺𝑅 converges to the reference value in the latter

part of the transient. The cylinder inputs for both the control strategies is shown in Figure

VII-6 and Figure VII-7. MPC retards the spark timing and advance the intake cam timing

heavily compared to the reference inputs during the tip-in maneuver.

Figure VII-3. Intake throttle mass flow rate trajectories for MPC and Open-loop control

Figure VII-4. Wastegate mass flow rate trajectories for MPC and Open-loop control

92
Figure VII-5. External cooled EGR mass flow rate trajectories for MPC and Open-loop

control shows that MPC deliberately stops EGR flow on the aggressive tip-in

Similar to the air-path actuators, the deviation of the MPC inputs from the

reference input is mainly to improve the turbocharger spool-up. Figure VII-8, Figure

VII-9 and Figure VII-10 shows the modeled volumetric efficiency, measured exhaust

temperature and measured turbocharger speed for both of the control strategies. The

advanced intake cam timing is clearly to improve the transient volumetric efficiency of

the cylinders whereas the retarded spark timing is to increase the exhaust temperature.

93
Figure VII-6. Spark timing trajectories for MPC and Open-loop control

Figure VII-7. Camshaft position timing trajectories for MPC and Open-loop control

94
Figure VII-8 ANN Modeled volumetric efficiency using MPC inputs and open-loop

inputs for tip-in maneuver

Figure VII-9 Measured turbine inlet temperature for MPC and Open-loop control shows

that MPC commands late spark timing to increase specific exhaust gas enthalpy for

turbocharger response

95
Figure VII-10 Turbocharger speed for MPC increases much more rapidly than open-loop

control

These multiple coordinated inputs result in the turbocharger speed increasing

significantly faster than the open-loop control. It is noteworthy, that once the IMEPn

tracking has reached steady state, the deviation between the MPC and the open-loop

inputs is minimal.

The most relevant constraint for the tip-in maneuver is the 𝐾𝐼 2 constraint. The

MPC advances the spark timing well beyond the reference value between 3 and 4

seconds. However, MPC also commands 𝑚̇𝐸𝐺𝑅 beyond the reference value as well during

this time period. The measured 𝐾𝐼 2 for cylinder 1 is shown in Figure VII-11. Due to the

spark timing advancement, the 𝐾𝐼 2 value increases well beyond the open-loop control

trajectory. However, due to the addition of extra 𝑚̇𝐸𝐺𝑅 into the engine, the constraint is

not violated.

96
Figure.

Figure VII-11 Measured 𝐾𝐼 2 for MPC does not exceed the limit of 1 despite advanced

spark timing during tip-in

7.2.2 Experimental test results 2900 RPM: Aggressive tip-out maneuver

Because the open-loop control strategy does not have active compressor surge handling

ability, only the MPC trajectories are shown. Figure VII-12 shows that there is a slight

instability in IMEPn after the tip-out.

97
Figure VII-12. Tip-out IMEPn response for MPC and Open-loop control strategy

The air path trajectories for the tip-out maneuver are shown in Figure VII-13. The MPC

commands the intake throttle mass flow rate to zero to reduce the intake manifold

pressure. MPC commands stoppage of EGR mass flow rate for combustion stability

concerns. Simultaneously the MPC commands an increase in waste-gate mass flow rate

and blow off valve mass flow rate to address the tip-out surge problem. Figure VII-15

shows that even though the MPC commands the blow off valve to open instantaneously,

the compressor operation actually crosses over into the surge region. This is mainly

because, the boosted engine operation before the tip-out is very close to the surge line. At

the instant of the tip-out the blow off valve command issued by the MPC cannot be

instantly fulfilled by the lower level controller due to bandwidth limitations. The true

compressor mass flow rate dynamics during the surge is also governed by very fast limit-

cycle like behavior which can be modeled using the classical Moore-Greitzer model

98
mentioned in [53]. However, addition of this model introduces dynamics which are

significantly faster than the MPC sample time to and should effectively be used in the

lower level control layer.

Apart from compressor surge avoidance, tip-out dilution from trapped external

EGR in the air-path is a concern as well. Since this EGR cannot be evacuated by any

other means other than by passage through the engine itself, the MPC can only control

the engine actuators to handle this problem. It is visible from Figure VII-14 that MPC

commands over-advanced spark timing at the instant of the tip-out to maintain

combustion stability. Although, internal residual gas fraction is not included in the

control model, the COV model indirectly captures the effect of excessive valve overlap

on COV which is the most likely reason for valve timing manipulation at the instant of

tip-out.

99
Figure VII-13 Air-path trajetories for aggressive tip-out maneuver

Figure VII-14 Cylinder actuator trajectories for aggressive tip-out maneuver

100
Figure VII-15 Compressor operation during the aggressive tip-out

101
 VIII. CONCLUSIONS AND RESEARCH CONTRIBUTIONS

Summary of research work

An evaluation of state-of-the-art control methods for turbocharged engine control

was performed in order to identify the lapses in research and potential for improvements

was identified. The synergies between various actuators on turbocharged SI engines is an

open research topic. The recent research work published for MPC based subsystem

control of turbocharged engines was examined in detail and it was quickly identified that

majority of research articles utilize linear system models which are either physics based

or of black-box nature and derived offline. Although this approach works well for air-

path control, is questionable when engine cylinder actuators are also included in the MPC

framework. Hence, in order to assess the error associated with the linearity approximation

a control oriented physics based model of the turbocharged engine air-path was

developed with ANN based sub-models for outputs and constraints. A study was

performed to evaluate the deviation of the non-linear model from the linearized model

over the prediction horizon which revealed that the linear approximation error is not

trivial for the highly non-linear and the most important outputs of the engine, IMEPn.

Extending this study further, a simple analysis was performed to evaluate the number of

re-linearizations required within the prediction horizon for a given required accuracy

level and operating point. The results of this analysis show that the model has varying

levels of linear approximation error depending on the operating condition at which the

analysis is done and the frequency of the reference inputs applied to the model.

102
 Under the knowledge that the engine model is non-linear and MPC strategies

require linearization of the model, avenues to speed up the linearization process have

been explored. Analytical differentiation has been widely utilized to reduce the number

of computations associated with model linearization. The resulting approach adopted has

been demonstrated to be at least 10 times faster in obtaining a linearized system model.

This is a significant advantage with cascaded benefits especially when the linearization is

done more frequently within the prediction horizon. Based on this approach, an NMPC

where linearization is done at every stage within the prediction horizon is formulated. A

preliminary co-simulation with 1D-Engine simulation has been performed and the results

have demonstrated the ability of the control system to coordinate air-path and engine

cylinder actuators simultaneously while tracking the desired IMEPn within 1 bar.

In order to facilitate experimental validation of the NMPC a lower level control

and turbocharger speed estimation layer was developed. Simple inversion of the orifice

flow model was performed with an additional adaptive filter to mitigate chattering of the

intake throttle control at very low pressure differential values. A fixed-point iteration

based turbocharger speed estimation algorithm was utilized to estimate the turbocharger

rotational speed using the production sensor set. A two micro-controller based rapid-

control-prototyping system was configured to facilitate real-time execution of the NMPC

algorithm on a computationally powerful micro-controller and the actuator control and

turbo speed estimation on a simpler machine. A sufficiently fast communication system

was established between the CAN controllers. Using this layout, real-time experimental

validation of the NMPC algorithm on a 2.0 L Turbocharged gasoline engine with cooled

103
external EGR was performed. In the absence of a production type engine control

algorithm for this engine configuration, the NMPC performance was evaluated against

the reference open loop control which is optimized for fuel economy. The NMPC

performs control actuator maneuvers similar to those observed using the GT-Power co-

simulation. The NMPC demonstrates the ability to manipulate Spark and cam timing to

improve engine breathing and exhaust gas enthalpy to maximize turbocharger response

while simultaneously controlling the air-path actuators. However, the surge control has

pitfalls due to the proximity of the compressor operation to the surge line and the

bandwidth limitations of the lower level control which were unaccounted for in the GT-

Power co-simulation.

Broader impacts and practical aspects

The algorithm outlined in this work is derived mainly from a physics based air-

path model coupled with ANN based engine model derived from steady state data. These

models were derived using 40 hours of steady state engine dynamometer test data. This

time period is negligible compared to the time required to acquire transient data to tune

black-box models of the engine. Furthermore, the physics based nature of the air-path

implies that changes in air-path geometry and turbocharger hardware can be quickly

implemented in the model without the need to re-identify parameters associated with air

path dynamics. These attributes of this approach imply that the algorithm can be ported to

a different engine system with minimal engine testing.

104
 Due to demonstrated ability of the algorithm to co-ordinate multiple actuators

simultaneously, calibration efforts required are minimal with only 15 constants which

require tuning, which is negligible compared to the large number of lookup tables and

maps associated with transient control of turbocharged engines on production ECUs.

Future work and potential for improvement

The NMPC proposed in this research is based on the solution to an MPC-QP

formulated based on the current state of the system, user defined objective function and

the reference inputs applied to the system. Although an MPC-QP is convex due to

linearized models used in the formulation, the original Non-linear optimization problem

can be non-convex. As a result, applying the sub-optimal control inputs derived from the

MPC-QP optimization may violate one or more constraints. This approach was used in

this research because 2 out of the 5 constraints on the mainly to impose physical

feasibility in the control model and can never be violated in reality. The remaining 3

operational constraints could possibly be violated with this approach. In order to find sub-

optimal solutions which satisfy the non-linear constraints, non-linear optimization

techniques like Sequential Quadratic Programming and Interior-point optimization

methods could be explored. However, these methods would require additional

computational resources.

The tuning of the NMPC is also an open question, wherein the duality between Q

and R implies that IMEPn tracking affects the cost associated with tracking of the

reference inputs and vice versa. Drive cycle or extended transient simulation could be

105
performed with different tuning values to study impact on fuel-economy. Improved

transient turbocharger response may not necessarily imply reduced fuel economy, as

consideration of interaction with transmission and shifting strategies is also a significant

factor in fuel economy.

106
 A. APPENDIX

A Control Algorithm for Low Pressure – EGR Systems using a Smith Predictor with

Intake Oxygen Sensor Feedback

Due to ever increasing stringency in emissions and fuel economy regulations

development of cost-effective and production viable concepts for engine efficiency

improvement is necessary. Utilization of cooled EGR in downsized turbocharged

gasoline engines has been shown to improve efficiency at many engine operating

conditions [1][2][3]. Previous research work in the domain of LP-cEGR control focuses

mainly on utilization of mean value models of air and EGR paths for determination of

EGR and air fraction [12]. An open-loop EGR valve mass flow model is utilized in these

methods [4][5]. The primary challenges associated with control of LP-cEGR systems are;

(1) mass flow modeling across EGR valve due to low-pressure differential across the

system [4][5], and (2) EGR control is difficult during transients due to long length of

EGR path [4][5][12]. Additionally, the methods described in [4] and [12] also utilize an

intake throttle valve upstream of the EGR mixing location, which increases pressure

differential across the EGR system. Additional throttle valves for EGR control are not

used for this research.

Universal Exhaust Gas Oxygen (UEGO) sensors are widely used in engine exhaust

systems for control and diagnostics. These sensors measure the concentration of oxygen

in the sampled gas. Previous research has demonstrated the feasibility of using UEGO

107
sensors to accurately measure the EGR fraction in the intake system [7][8]. The potential

of these sensors to provide real-time EGR fraction feedback control is explored in this

research. Water condensation from exhaust [1] can severely damage the high temperature

sensing element in these sensors. Also, in order to ensure proper EGR and Air mixing at

the sensor location, the sensor location has to be adequately downstream of the mixing

location. However, moving the sensor farther downstream of the EGR valve increases the

EGR valve to sensor time delay. This severely restricts the calibrated ‘aggressiveness’ of

the feedback control system. In systems with dead time delays, Smith Predictor based

feedback control is a widely used method for process control [13][6]. Smith Predictor

based control is a predictive method of process control where the measured output of the

process is delayed by a known time duration. Fluid flow through a pipe and engine

cylinders are common examples of processes with dead time delay.

The paper begins with description of experimental setup including the engine and control

system hardware. As required by the Smith Predictor method, models are defined for the

EGR and air path. These consist of open-loop modeling of EGR valve mass flow and the

transport delay across the EGR and air path. These models are validated in isolation to

evaluate their accuracy. Finally, the complete control structure is tested at steady state

engine operating conditions to evaluate performance and stability.

A downsized turbocharged inline 4 gasoline spark-ignition (SI) engine with

specifications described in Table 1 is used for testing. An AC engine dynamometer

controls engine speed and/or loading for all experiments. A set of production-intent

sensors is installed on the engine for basic engine functionality (air, fueling, ignition,

108
valve timing and turbo control). A low-pressure and cooled EGR (LP-cEGR)

configuration is retrofitted on the engine. Exhaust gases are extracted downstream of the

turbine and upstream of the catalyst. EGR passes through a cooler and is delivered to the

intake air-path system upstream of the compressor. The EGR cooler is a tube-core type

chosen for low pressure loss characteristics. Twin liquid-to-air intercoolers have been

used to allow high boost/load capability in the dynamometer cell, where air/air heat

exchanger effectiveness is low.

Engine type Turbocharged Inline 4 cylinder

SI
Displaced volume 1998 cc

Bore x Stroke 86 x 86 mm

Compression ratio 9.5:1

Valve train DOHC Continuously variable

valve timing with 4 valves per

Fuel system Direct Injection
cylinder
EGR system Low-pressure cooled EGR

Base engine controller BOSCH SI Engine controller

Rapid prototype ETAS ES910

controller

109
The overall EGR control system architecture is outlined in Figure 2, and a description of

the signals is provided in Table 2. A production-intent engine controller is modified to

include software hooks on specific engine control parameters. An ETAS ES910 Rapid

Prototype Controller (RPC) is used in conjunction with the base engine controller using

CAN Communication Protocol (CCP) bypass communication. This is done to establish

data interface between the base engine controller and the rapid prototype controller. A

variant of an oxygen sensor is also installed post compressor (Figure 1.). This sensor is

similar in operation to UEGO sensors, however it has been developed and optimized to

be operated specifically in the intake environment, where temperature range and gas

composition is different from the exhaust. EGR fraction is calculated based on change in

intake oxygen concentration caused by EGR dilution [7][8]. Hereafter, the intake oxygen

sensor is referenced as the ‘EGR sensor’ used for feedback control.

Production Sensors
and actuators

Base
engine
controller
……

Engine
Additional Sensors

CCP bypass
and actuators

RPC EGR
Controller

Figure A-1 Control system setup

110
 𝑚̇𝐴𝑖𝑟 Fresh air mass flow measured before EGR mixing
𝑃𝐶𝑜𝑚𝑝−𝐼𝑛 Compressor inlet pressure measured at same
𝑃𝐶𝑜𝑚𝑝−𝑂𝑢𝑡 Boost pressure measured post intercooler
𝑇𝐴𝑖𝑟 location as 𝑚̇𝐴𝑖𝑟
Inlet air temperature measured at 𝑚̇𝐴𝑖𝑟 location
𝑁𝐸𝑛𝑔 Engine speed [rpm]
𝑇𝐸𝐺𝑅 EGR cooler outlet exhaust temperature
𝑥𝑂2𝑆𝑒𝑛𝑠 Mole fraction of Oxygen measured by EGR sensor

𝜃𝐸𝐺𝑅−𝑉𝑎𝑙𝑣𝑒 Commanded EGR valve angle

EGR Valve Modeling for Control

Mass flow rate across the EGR valve is modeled using an isothermal orifice flow model

in equation (76). This model is specifically used for exhaust gases at lower temperatures

83.

𝐶𝐷 .𝐴𝑉𝑎𝑙𝑣𝑒 .𝑃𝐼𝑛 2𝑃𝑂𝑢𝑡 𝑃𝑂𝑢𝑡 1/2

𝑚̇𝐸𝐺𝑅 = { [1 − ]} 83
𝑃𝐼𝑛 𝑃𝐼𝑛
√𝑅. 𝑇𝐸𝐺𝑅

⁡𝐶𝐷 is the discharge coefficient and 𝐴𝑉𝑎𝑙𝑣𝑒 is the valve open area. 𝑃𝐼𝑛 is the valve inlet

pressure. To determine pressure at the inlet of the EGR valve, the pressure at the inlet of

the EGR cooler and the pressure drop across the EGR cooler needs to be known.

Calculation of pressure drop across the EGR cooler requires prior knowledge of mass

flow across the EGR system. For the sake of simplicity, this research considers EGR

valve inlet pressure to be the same as EGR cooler inlet pressure. A physics based model

is utilized to predict this pressure as a function of engine operating conditions [68]. 𝑃𝑂𝑢𝑡

111
is the valve outlet pressure which is considered to be the same as 𝑃𝐶𝑜𝑚𝑝−𝐼𝑛 . 𝑅 is the

specific gas constant for stoichiometric exhaust gas. LP-cEGR systems are required to

operate at a much lower pressure differentials as compared to HP-EGR systems. The

pressure ratio correction is very sensitive to minor inaccuracies in the pressure ratio at

such low pressure differences. In addition, the pressure at the inlet of the EGR valve

pulsates[4][5], which influences the flow through the EGR system. The model shown in

(83) holds strictly true only for steady flow conditions. Pressure pulsations reduce the

accuracy of the model as the flow is unsteady at the valve and averaged pressures are

used as inputs to the model.

In this research the discharge coefficient used in the model has been characterized as a

function of engine speed and EGR valve angle. This is done to partially capture the effect

of the primary frequency of pressure pulsations on the discharge coefficient since the

primary frequency of exhaust pressure pulsations is directly proportional to the engine

speed [11]. The discharge coefficients were obtained by running an optimization routine

which minimizes the error between the modeled mass flow and the measured mass flow.

The measured mass flow is derived from experimental sweeps of EGR valve position

between fully closed and fully open at different engine operating conditions. The engine

load was swept from 3 to 11 Bar BMEP and theengine speed was swept from 1200 to

2700 RPM. The final result of this optimization routine is the two-dimensional discharge

coefficient map shown in Figure A-2. Typical maximum values of discharge coefficient

for butterfly type valves are close to 0.6 [11]. The values obtained for this research were

signficantly lower (0.15) mainly due to consideration of EGR cooler inlet pressure as

112
valve inlet pressure. A minor inaccuracy in the mapping between EGR valve angle and

open area has also influenced the discharge coefficient.

Modeled and measured mass flow rates along with the +/-10% error lines are

shown in Figure A-3. The majority of operating points lie within the +/-10% range.

Engine speed has been used to capture the effect of exhaust pressure pulsations. The

primary frequency of exhaust pressure pulsations is a linear function of engine speed and

number of cylinders/exhaust blowdown events per cycle. This has been done to partially

characterize the effect of exhaust pressure pulsation frequency on the EGR mass flow.

Effects of peak amplitudes and pulsation shape may also have an impact, but are outside

of the scope of this research

0.15
Discharge coefficient [-]

0.1

0.1

0.05

0.05

80
70
0
60
2700
50 2500
40 2300
2000
30 1800
EGR valve angle [deg] 20 1500 Engine speed [rpm]
1200

Figure A-2 Dimensional Discharge coefficient map obtained from optimization routine

113
Figure A-3 Measured vs modeled EGR valve mass flow rate

Transport delay model

The dead-time delay between EGR valve and the EGR sensor consists of two

components. The first is transport delay from the EGR mixing location to EGR sensor.

The second component is sensor response time, consisting of transport from exterior of

sensor to sensing element and electrical signal delay. The sensor response time is

approximated as a linear function of flow velocity in the control volume in which the

sensor is mounted. A simplified equation based on the assumption of plug flow is derived

using the steady state continuity equation for mass flow rate through a control volume.

114
This method is utilized to estimate the transport delay from the entry to the exit of the

control volume. The equation used to estimate the transport delay 𝜏𝑖 ⁡across a generic

control volume ‘i’ in the EGR system is given by equation (84).

𝐿 .𝑃 .𝐴
𝜏𝑖 = 𝑚𝑖̇ .𝑅.𝑇
𝑖 𝑖
84
𝑖 𝑖

In this equation 𝐿𝑖 and 𝐴𝑖 are the length and flow area respectively. 𝑚̇𝑖 , 𝑃𝑖 and 𝑇𝑖 are the

mass flow rate, pressure and temperature in the control volume respectively. Although

the equation is derived for strictly steady state conditions, it is still used for the quasi

steady state operating conditions tested in this research where only EGR valve transients

were performed.

EGR fraction models

Implementation of the Smith Predictor methodology requires definition of the

EGR fraction model with no transport delay and the EGR fraction model with modeled

transport delay. These models are described as per the schematic in Figure A-4.

Figure A-4 Modeled low pressure EGR path

115
 The EGR fraction model with no transport delay is used to predict the EGR

fraction at mixing point A. Under the assumption that the mixing is instantaneous and

perfect, the EGR fraction at A, 𝐸𝐺𝑅𝑓𝐴 is given by equation (85).

𝑚̇𝐸𝐺𝑅
𝐸𝐺𝑅𝑓𝐴 = 𝑚̇ 85
𝐴𝑖𝑟 +𝑚̇𝐸𝐺𝑅

The EGR fraction at location B and time, 𝑡, 𝐸𝐺𝑅𝑓𝐵 (𝑡)⁡which is also the EGR

sensor location is the transport delayed value of the EGR fraction at A. The modeled

EGR fraction at B is given by equation (86).

𝐸𝐺𝑅𝑓𝐵 (𝑡) = 𝐸𝐺𝑅𝑓𝐴 (𝑡 − 𝜏(𝑡)) 86

Where 𝜏(𝑡)⁡is the modeled transport delay at time 𝑡 from location A to location B,

and is determined using equation (87).

𝜏(𝑡) = 𝜏𝐴𝑡𝑜𝐶 (𝑡) + 𝜏𝐶 (𝑡) + 𝜏𝐶𝑡𝑜𝐵 (𝑡) 87

Where 𝜏𝐴𝑡𝑜𝐶 (𝑡) and 𝜏𝐶𝑡𝑜𝐵 (𝑡) are the modeled transport delays from A to the

compressor and compressor to B respectively. 𝜏𝐶 (𝑡) is the transport delay within the

compressor which has been assumed to be zero. These delays are calculated separately

because the pressure and temperature within the control volumes from A to the

compressor and from the compressor to B are different.

116
EGR fraction measurement using Oxygen sensing

The EGR mole fraction, 𝐸𝐺𝑅𝑓𝐵 𝑆𝑒𝑛𝑠 , is the measurement performed by the EGR

sensor. It is calculated using the measured mole fraction of oxygen ⁡𝑥𝑂2 𝑆𝑒𝑛𝑠 ⁡with equation

(88).
𝑥𝑂2 −⁡𝑥𝑂2
𝐴𝑚𝑏 𝑆𝑒𝑛𝑠
𝐸𝐺𝑅𝑓𝐵 𝑆𝑒𝑛𝑠 = 88
𝑥𝑂2 −⁡𝑥𝑂2
𝐴𝑚𝑏 𝐸𝑥ℎ

𝑥𝑂2 𝐴𝑚𝑏 and ⁡𝑥𝑂2 𝐸𝑥ℎ are the mole fractions of oxygen in fresh air and the exhaust gases

respectively. In order to derive the EGR mass fraction, the following assumptions are

made:

 Effect of ambient humidity on 𝑥𝑂2 𝐴𝑚𝑏 is ignored

 ⁡𝑥𝑂2 𝐸𝑥ℎ is negligible due to mostly stoichiometric operation

 The difference between EGR mass fraction and EGR mole fraction is considered

negligible for control simplicity

Feedback control with Smith Predictor

A classical Smith Predictor was implemented as a dead time delay compensator. Previous

research shows that Smith Predictor based classical feedback control for EGR systems in

Large Diesel Engines improves control stability [61]. The application of the Smith

predictor is visible in the EGR fraction controller shown in Figure A-5. Linear,

continuous approximations of subsystems have been shown. Closed loop feedback

control of EGR fraction is performed using a PID controller.

117
Figure A-5. Closed loop EGR fraction controller with Smith predictor

𝑃𝑃𝑙𝑎𝑛𝑡𝐵 (𝑠) ≈ 𝑃𝑃𝑙𝑎𝑛𝑡𝐴 (𝑠). 𝑒 −𝜏𝑃𝑙𝑎𝑛𝑡(𝑡)𝑠 89

𝑃𝑃𝑟𝑒𝑑𝐴 (𝑠) = 𝑃𝑀𝑜𝑑𝑒𝑙𝐴 (𝑠) + (𝑃𝑃𝑙𝑎𝑛𝑡 𝐴 . 𝑒 −𝜏𝑃𝑙𝑎𝑛𝑡 (𝑡)𝑠 − 𝑃𝑀𝑜𝑑𝑒𝑙𝐴 (𝑠). 𝑒 −𝜏(𝑡)𝑠 ) 90

The primary objective of this architecture is to eliminate the delay dynamics from the

output. The EGR Path + Sensor subsystem system has the output 𝐸𝐺𝑅𝑓𝐵 𝑆𝑒𝑛𝑠 which is

inherently delayed by the physical transport delay 𝜏𝑃𝑙𝑎𝑛𝑡 (𝑡) as shown in equation (89)

𝑃𝑃𝑙𝑎𝑛𝑡𝐴 (𝑠) is the EGR Path + Sensor subsystem without the transport delay. The output

of augmented system approaches that of the Mixing point EGR fraction model as the

term in the parenthesis in equation (90) tends to 0 i.e. as the output of 𝑃𝑀𝑜𝑑𝑒𝑙 𝐴 (𝑠) tends

to 𝑃𝑃𝑙𝑎𝑛𝑡 𝐴 (𝑠) and 𝜏(𝑡) to⁡𝜏𝑃𝑙𝑎𝑛𝑡 (𝑡). This emphasizes the importance of accuracy of the

EGR flow model and transport delay model. Effectively, the output 𝐸𝐺𝑅𝑓𝐴 𝑃𝑟𝑒𝑑 of the

118
augmented system⁡𝑃𝑃𝑟𝑒𝑑 𝐴 (𝑠) has minimal delay dynamics in it. This reconstructed output

along with the desired EGR fraction command 𝐸𝐺𝑅𝑓𝐵 𝐶𝑜𝑚𝑚 is used to generate the error

upon which the PID controller operates. Therefore, instability and oscillation induced by

dead time delays are mitigated.

Experimental test results

The benefits of LP-cEGR are seen at different engine operating conditions due to

different mechanisms [36]. This research mainly focuses on LP-C-EGR control at low to

mid load. This is done partially due to the generally longer transport delay 𝜏(𝑡)⁡associated

with these load conditions. Longer transport delays can exist for high load and very low

engine speeds but for this research these conditions were not considered.

Transport delay model

The transport delay model was tested experimentally by performing step changes in EGR

valve position at quasi steady engine operating conditions. These conditions were

established by operating the engine at fixed actuator positions/commands; such as the

intake throttle valve, valve timing, waste-gate and blow-off valve position, etc. The

number of engine cycles elapsed between the EGR valve step and the effect of the step at

the EGR sensor was compared with the delay calculated by the transport delay model.

This is done at various engine operating conditions to account for varying engine cycle

119
time. The majority of test points lie within a +/- 1 engine cycle band, as shown in Error!

Reference source not found..

Figure A-6 Modeled vs measured transport delay

EGR fraction models

Recall that 𝐸𝐺𝑅𝑓𝐵 is the transport delayed version of 𝐸𝐺𝑅𝑓𝐴 using the transport delay

model. In order to evaluate accuracy of the EGR fraction models EGR valve steps were

performed, and modeled and measured EGR fractions at the EGR sensor location are

compared Figure A-7 to Figure A-9. show this comparison at various engine operating

conditions for different EGR valve position steps. At all of the operating conditions, the

fast change in 𝐸𝐺𝑅𝑓𝐵 and 𝐸𝐺𝑅𝑓𝐵 𝑆𝑒𝑛𝑠 caused by the EGR valve position step are within

+/- 1 engine cycle. The steady state difference between 𝐸𝐺𝑅𝑓𝐵 and 𝐸𝐺𝑅𝑓𝐵 𝑆𝑒𝑛𝑠 is highest

120
at 2100 RPM and 2.5 Bar BMEP. This is attributed to the inaccuracy of the EGR mass

flow model at near unity pressure ratio.

Figure A-7. Comparison between modeled and measured EGR fraction at EGR sensor

location; engine operated at 1700 RPM 3 Bar BMEP

121
Figure A-8. Comparison between modeled and measured EGR fraction at EGR sensor

location; engine operated at 2100 RPM 5 Bar BMEP

122
Figure A-9. Comparison between modeled and measured EGR fraction at EGR sensor

location; engine operated at 2100 RPM 2.5 Bar BMEP

Feedback control with Smith Predictor

In order to test the feedback control architecture outlined in Figure A-5, step inputs are

commanded for desired EGR fraction. These tests are performed using conventional PID

control in Figure 11, Figure 13 and Figure 15. Step response of EGR fraction controller

with conventional PID controller at 2000RPM 9 Bar BMEP as well as Smith Predictor

based control as seen in Figure 12, Figure 14 and Figure 16. Step response of EGR

fraction controller with Smith Predictor at 2000RPM 9 Bar BMEP The controller gains

Kp, Ki and Kd were determined heuristically as 130, 6 and 0.01 respectively by step

123
response analysis. They were kept constant for all of the tests. In the case of the

conventional PID control there are two main observations:

 The EGR valve position overshoots the position required to reach the desired

EGR fraction

 The EGR valve position appears to be 90 degrees out of phase with the sensed

EGR fraction, thereby inducing oscillation in the EGR fraction

Oscillations can be mitigated with appropriate controller tuning. However, the controller

gains would also have to be scheduled as per engine operating condition as the flow

within the EGR and air path is mainly dependent on engine operating condition.

Figure A-10. Step response of EGR fraction controller with conventional PID controller

at 1500RPM 7 Bar BMEP

124
Figure A-11. Step response of EGR fraction controller with Smith Predictor at 1500RPM

7 Bar BMEP

In case of the Smith Predictor based controller the following observations are made:

 Oscillations related to phase lag between EGR valve and sensor are greatly

reduced (compared to PID)

 Overshoots in EGR fraction are greatly reduced (compared to PID)

 Rise time and fall time (time duration from 10% to 90% of the setpoint value) are

increased over conventional PID

125
There are small overshoots in the EGR valve position, but they are small enough in

magnitude and time to not affect EGR fraction significantly, as evident in Figure A-13 at

3 and 27 seconds. There are some steady state oscillations in EGR fraction as well. These

were determined to be a result of a ‘sticky’ EGR valve and controller combination. The

increase in rise and fall time are mainly due to the difference between modeled and

measured EGR fractions. The comparison between the performance criteria of these two

control strategies is summarized in Table A-1.

Figure A-12. Step response of EGR fraction controller with conventional PID controller

at 1300RPM 4 Bar BMEP

126
Figure A-13. Step response of EGR fraction controller with Smith Predictor

127
Figure A-14. Step response of EGR fraction controller with conventional PID controller

at 2000RPM 9 Bar BMEP

Figure A-15. Step response of EGR fraction controller with Smith Predictor at 2000RPM

9 Bar BMEP

128
Table A-1 Conventional PID control vs Smith Predictor Control performance comparison

Performance Conventional PID Smith Predictor

criteria Control Control

Average overshoot 36.3 12

[%]
Average rise time 0.823 1.23
[sec]
Average fall time 0.383 1.13
[sec]
Steady state behavior Oscillating Stable

Summary/Conclusions

A closed loop LP-cEGR feedback controller with Smith predictor based dead time delay

compensator was implemented and tested. Open loop models for EGR fraction at the

mixing and EGR sensor locations along with a transport delay model were developed and

129
validated. EGR valve step changes were performed to evaluate step response of a

conventional PID controller with that of Smith Predictor based control.

The Smith Predictor control showed significant reduction in average overshoot as well as

improved steady state stability. However, due to mismatch between EGR sensor feedback

and modeled EGR fraction, rise time and fall time increased as compared to PID control.

Accurate open loop modeling of EGR mass flow in LP-cEGR systems is still an area

where development is required. The conventional orifice flow model begins to lose

accuracy at low pressure differentials and with pressure pulsations. These conditions

require additional calibration and tuning parameters to be usable for control.

130
 Nonlinear Model Predictive Control of Dual Loop - Exhaust Gas Recirculation in a

Turbocharged Spark Ignited engine

The external Exhaust Gas Recirculation (EGR) technology is widely used on

Compression Ignition (CI) engines for its ability to reduce NOx emission [69]. This

technology was introduced to Spark Ignition (SI) engines recently due to its ability to

mitigate knock and reduce pumping loss[11]. However, SI engines require much more

precise EGR concentration control than the CI engines. The employment of High Pressure

(HP) and Low Pressure (LP) dual loop (DL) EGR provides more degrees-of-freedom for

EGR concentration control. The LP-EGR loop has a characteristic transport delay much

longer than the HP-EGR loop, which introduces additional difficulty in air fraction control

in the intake manifold. Conversely, exhaust enthalpy exchange through the turbine is

reduced in the case of HP-EGR thereby reducing turbine power and hence adversely

affecting compressor mass flow rate. It is desirable to coordinate the dual loop EGR

systems to achieve given EGR concentration target and maximize turbocharger efficiency.

In this work, a Nonlinear MPC (NMPC) is proposed to control the intake manifold pressure

(MAP) and EGR concentration of a turbocharged SI engine with DL-EGR, by

manipulating the throttle, EGR valves of HP and LP loops. For a given preview horizon of

tacking references, the NMPC minimizes tracking error and control effort subject to the

actuator limit and states constraints.

The control of turbocharged gasoline engine air handling system with EGR is a

Multi-Input Multi-Output (MIMO) problem with nonlinear and coupled dynamics. NMPC

has been utilized in [93], to track torque for a turbocharged SI engine where the throttle

131
and wastegate are considered as the control inputs. External EGR control has not been

included in the aforementioned research. While references for the DL-EGR control of SI

engine are limited, the control of the diesel engine air system with dual loop EGR has been

extensively discussed in literature [69] and [94], where the authors successfully

implemented a switched linear MPC with feed-forward and feedback controllers for the

multi-input multi-output nonlinear DL EGR and VGT system. In these papers, multiple

linear models were developed over engine speed and fueling operating regions and a

switched linear MPC approach with quadratic cost function was employed to calculate

control actions.

The EGR control strategy developed for CI engines may not be suitable for SI

engines. The LP EGR transport dynamics considered in previous CI engine research are

highly simplified and usually similar to a first order delay. This approximation may be

tolerable for CI engine operation as the ignition process is more robust against EGR

dilution. Such an assumption for a SI engine can potentially lead to undesirable operation

like misfire and partial combustion due to over dilution of EGR. Hence, it is necessary to

improve the accuracy of transport delay dynamics in the control oriented model of the

airpath as shown in Figure A-16, wherein a NMPC has been utilized with a segmented

boost manifold based transport delay model integrated into the airpath model. This research

employs a similar approach in which a high order transport delay model is utilized to

improve the prediction of LP-EGR fraction from the location of its delivery to the engine

cylinders.

132
Control oriented air path model

The individual components of the air path have been modeled separately and

ultimately integrated to form a control oriented air path as depicted in Error! Reference

source not found..

Figure A-16. Schematic of Engine and Air path model.

Here, 𝑃, 𝑇, 𝐹, 𝑚̇, 𝜔, and 𝑉 are pressure, temperature, mass fraction, mass flow rate,

rotational speed and volume respectively. Subscripts𝐸𝑥ℎ, 𝐵𝑠𝑡, 𝐼𝑚, 𝐻𝑃𝐸, 𝐿𝑃𝐸, 𝐼𝑇𝑉,

⁡𝐶𝑜𝑚𝑝, 𝑇𝑢𝑟𝑏, 𝐼⁡⁡and⁡𝐸 are the exhaust manifold, boost manifold (between compressor and

throttle), intake manifold, high-pressure EGR, low-pressure EGR, intake throttle,

compressor, turbine, ambient compressor inlet and post-turbine exhaust respectively. The

133
following section is dedicated to description of the component level models used in the air

path model.

Gas transport dynamics in boost manifold

The boost manifold is typically the largest control volume in a turbocharged engine air path

and hence consideration of the gas transport dynamics in the boost manifold is critical for

transient control of EGR dilution at the cylinders. This manifold has been traditionally

considered as a lumped volume in the turbocharged diesel engine air path. However, in

typical turbocharged automotive engines the boost manifold is characterized more by thin

long pipes rather than plenums. Therefore, a lumped volume based assumption for the

boost manifold is done at the expense of accuracy. Experimental testing has confirmed the

above mentioned phenomenon as shown in Figure 4 where a step input in EGR valve

position results in a dead time delayed and mixing influenced response from the EGR

measurement sensor installed at the throttle valve. The mixing influence cannot be

attributed purely to mixing in the manifold but also to the filling and emptying dynamics

of the sensor itself.

134
Figure A-17. Step response of oxygen sensor based EGR measurement installed at

compressor outlet location

The consequence of excessive EGR dilution in diesel engines typically results in increased

smoke and soot emission. However, with excessive EGR dilution, gasoline spark ignited

engines can suffer from very high cyclic variability in torque production. Furthermore, the

three way catalyst can encounter premature damage due to misfires caused by excessive

dilution. The gas transport model chosen here (91) is based on conservation of mass and

the ideal gas law in a control volume.

𝑅𝑇𝑖 91
𝐹𝑖̇ = (𝐹 𝑚̇ − 𝐹𝑖 𝑚̇𝑂𝑢𝑡 )
𝑃𝑖 𝑉𝑖 𝐼𝑛 𝐼𝑛

where, 𝐹𝑖 is the EGR fraction in the 𝑖 𝑡ℎ control volume 𝑉𝑖 . The number of control volumes

utilized to discretize the boost manifold is the same as the number of fraction states 𝐹𝑖 used

in the transport model. Hence, the order of the transport model is equal to the number of

sections used to discretize the control volume.

135
Figure A-18. Comparison of step response of transport delay models with true dead time

delay response

Higher order models show increasing accuracy and similarity to the ‘true’ dead time

delayed response (if pure pipe plug flow is considered) as shown in Figure A-18. A 5th

order transport delay model was chosen for this research as it has significantly higher

accuracy than the 1st order model which is a single lumped volume for the entire boost

manifold. Moreover, higher order models would have diminishing improvement in

accuracy over the computational burden of additional state dynamics.

136
State space model and control problem formulation

Using the component level models described in the previous section, iso-thermal

pressure dynamics and turbocharger rotor dynamics described in [14] the discrete time

state-space model is derived as given below. Since the goal of this research was to evaluate

control of mass flow rates of the EGR valves, temperature dynamics were ignored to keep

system dimensionality low and allow for more transport delay states.

𝑥(𝑘 + 1) = 𝑓(𝑥(𝑘), 𝑢(𝑘)) 92

𝑦 = 𝐶𝑥 = [𝑃𝐼𝑚 ⁡𝐹𝐼𝑚 ⁡]𝑇 93

𝑥 = [𝑃𝐼𝑚 ⁡𝑃𝐵𝑠𝑡 ⁡𝜔 𝑇𝑢𝑟𝑏𝑜 ⁡𝐹1 ⁡𝐹2 ⁡𝐹3 ⁡𝐹4 𝐹5 ⁡𝐹𝐼𝑚 ⁡𝑚̇𝐿𝑃𝐸 ]𝑇 94

𝑢 = [𝑚̇𝐼𝑇𝑉 ⁡𝑚̇𝐻𝑃𝐸 ⁡𝑚̈𝐿𝑃𝐸 ]𝑇 95

The output vector 𝑦⁡consists of the intake manifold pressure 𝑃𝐼𝑚 and intake

manifold EGR%⁡𝐹𝐼𝑚 . For gasoline SI engines the intake manifold pressure and EGR%

strongly associated with the engine load and these references are assumed to be available

from an external source (eg. Torque management system, supervisory control). The

remaining states 𝑃𝐵𝑠𝑡 , 𝜔 𝑇𝑢𝑟𝑏𝑜 , 𝐹𝑖 and 𝑚̇𝐿𝑃𝐸 are the compressor outlet pressure,

turbocharger rotational speed, the transport delay states and the LP-EGR mass flow rate

respectively. The input vector 𝑢 consists of the intake throttle and high pressure mass flow

rate 𝑚̇𝐼𝑇𝑉 and 𝑚̇𝐻𝑃𝐸 . The LP-EGR mass flow rate is controlled using input 𝑚̈𝐿𝑃𝐸 , as this

137
allows tuning of the rate of change of the LP-EGR mass flow rate. The model is also

normalized to mitigate potential numerical issues during execution.

The objective of the control algorithm is to minimize the squared error between a reference

value 𝑦𝑅𝑒𝑓 (𝑘) and the system output 𝑦(𝑘) for the preview horizon. Hence, the optimal

control problem has been designed to be an output tracking controller with the cost function

defined as follows.

𝐽(𝑥(𝑘), 𝑈(𝑘)) = ⁡𝑦̅(𝑘 + 𝑁)𝑇 𝐻𝑦̅(𝑘 + 𝑁) + ∑𝑘+𝑁−1

𝑖=𝑘 ⁡𝑦̅(𝑖)𝑇 𝑄𝑦̅(𝑖) + 𝑢(𝑖)𝑇 𝑅𝑢(𝑖) 96

where,

𝑦̅(𝑘) = ⁡𝑦(𝑘) − 𝑦𝑅𝑒𝑓 (𝑘) 97

3×3
ℝ2×2
>0 , 𝑅 ∈ ℝ>0

𝑈(𝑘) = [𝑢(𝑘), 𝑢(𝑘 + 1), … 𝑢(𝑘 + 𝑁 − 1)]𝑇 98

𝑄 and⁡𝐻 are the tuning matrices for the penalty on output tracking error within the

prediction horizon and the terminal output error. The tuning of these matrices was done to

place more emphasis on the tracking of 𝐹𝐼𝑚 as the consequence of poor tracking of this

parameter can result in misfires. 𝑅⁡is the tuning matrix for penalty on magnitude of input

𝑢. The non-linear optimization problem is hence formulated as follows.

min
𝐽(𝑥(𝑘), 𝑈(𝑘))
𝑈(𝑘)

𝑚̇𝐼𝑇𝑉 (𝑖) ≥ 0,

𝑚̇𝐻𝑃𝐸 (𝑖) ≥ 0,

𝑃𝐵𝑠𝑡 (𝑖) − 𝑃𝐼𝑚 (𝑖) ≥ 0,

138
 𝜔 𝑇𝑢𝑟𝑏𝑜 (𝑖) ≥ 0.

where, 𝑖 = 𝑘, 𝑘 + 1. . , 𝑘 + 𝑁 − 1

Figure A-19 Closed-loop control structure of DL-EGR NMPC

All of the constraints are imposed to maintain physical feasibility in the system.

The turbocharger rotational speed has the longest time constant in the system (1.1 seconds)

Hence, for a sample time of 0.1 seconds the horizon length 𝑁 is 11. The numerical values

chosen in the tuning matrices are chosen heuristically with more emphasis on the EGR%

control. Fig. 6 shows the overall schematic representation of Simulation results and

discussion the NMPC control structure. The NMPC problem was simulated using ACADO

toolkit [95].

139
 The simulation was performed to study the ability of NMPC to handle the following

aspects of the system:

 Reference tracking of 𝑃𝐼𝑚 and 𝐹𝐼𝑚

 LP-EGR transport delay in the boost manifold

 Interaction between EGR loops and turbocharger

 Control of split ratio between HP-EGR and LP-EGR

The intake manifold pressure reference trajectory was deliberately chosen to have

boosted and un-boosted values so that the interaction between the EGR circuits and the

turbocharger performance could be studied. Arbitrary time varying references were chosen

for EGR%. The simulation was performed for three engine speeds, 1000, 2000 and 3000

RPM with the same reference trajectories. The speeds were chosen because external cooled

EGR has maximum benefit in knock mitigation at low speed high load conditions. In the

following section simulation trajectories are shown for 2000 RPM.

High-pressure EGR control only

It is evident from Figure A-20 and Figure A-21 that in this case the system is able

to track the reference of the intake manifold mass better than the low pressure only and

dual loop cases. However, the intake manifold pressure reference tracking under a

reference value above atmospheric pressure is the poorest in this case as seen between 5

and 7 seconds. This is because of the reduction of mass flow rate through the turbine caused

by the high-pressure EGR system. The HP-EGR circuit effectively behaves like a waste-

gate and hence results in reduced turbine power. As per the perception of the driver, this

140
would seem as poor engine torque response. The controller overshoots the intake throttle

and HP-EGR mass flow rates whenever there is a sudden change in reference. Another

thing to note is that the boost pressure always stays above or equal to the intake manifold

pressure implying that 𝑚̇𝐼𝑇𝑉 is physically feasible.

Figure A-20 Pressure states and intake manifold EGR dilution trajectories for HP-EGR

control only.

141
Figure A-21 Input trajectories for HP-EGR control only.

142
Low-pressure EGR control only

The most significant difference between this case and the HP-EGR case is that the intake

manifold pressure reference tracking is significantly better as seen in Figure A-22 and

Figure A-23. as all of the exhaust mass flow passes through the turbine resulting in higher

turbine power to drive the compressor to higher boost levels. However, the reference

tracking for EGR mass in the intake manifold is worse with oscillation at the EGR%

reference changes at 5 and 8 seconds. Due to the transport delay model being

incorporated into the NMPC, there is also some lead action visible at 0.5 seconds on the

LP-EGR mass flow rate to meet the EGR% reference change at 1 second. The LP-EGR

mass flow rate has an immediate effect on only the EGR% in the first section of the boost

manifold which is far upstream of the intake manifold. The dilution dynamics in every

consequent air-path section diminishes the effect of this input on the intake manifold

EGR%. Hence, the tuning weight in matrix R associated with the rate of change of LP-

EGR mass flow rate had to be modified to reduce oscillations in the LP-EGR mass flow

rate.

143
Figure A-22 Pressure states and intake manifold EGR dilution trajectories for LP-EGR

control only.

144
Figure A-23 Input trajectories for LP-EGR control only.

145
Dual-loop EGR control

In this case the intake manifold pressure and EGR mass reference tracking appear

to be reasonably better compared to the LP-EGR only case as shown in Figure 9a. and 9b.

Since the engine is operated throttled up to 5 seconds the NMPC commands only HP-

EGR up to 4 seconds. At the transition from throttled to boosted operation LP-EGR flow

is initiated to maximize turbine power. As the intake manifold pressure level increases

due to turbo spool-up, a sharp HP-EGR mass flow rate spike is commanded at 5 and 8

seconds to increase intake manifold pressure further while maintaining the desired

EGR% reference. The penalty on rate of change of LP-EGR mass flow rate was also

relaxed compared to the LP- EGR only case. It is noteworthy that the commanded LP-

EGR mass flow rate shows lesser oscillations despite the relaxed penalty indicating that

NMPC effectively uses HP-EGR to ‘supplement’ the LP-EGR during boosted operation.

146
Figure A-24 Pressure states and intake manifold EGR dilution trajectories for Dual-loop

EGR control.

147
Figure A-25 Input trajectories for Dual-loop EGR control.

148
Additional comparison between different cases

The intake manifold EGR% for different cases at 2000 RPM is shown in Figure

A-25. Even though the HP-EGR case settles to the reference at 5 seconds, there is an

error associated with the first order filling dynamics of the intake manifold visible

between 4.9 and 5 seconds. This error is diminished for the LP-EGR and DL-EGR case

as the primary EGR actuator in these cases is the LP-EGR mass flow rate which is farther

downstream from the intake manifold.

Figure A-26 Transient intake manifold EGR dilution for multiple EGR architectures

The performance of the NMPC for dual and single loop EGR architectures has been

summarized in TABLE A-2. At 1000 and 2000 RPM the DL-EGR shows reduced RMSE

149
and Max errors compared to the single loop architectures. However, at 3000 RPM the

DL-EGR showed poorer performance than the single loop systems. This can be attributed

to non-ideal tuning for the output tracking and input penalties

TABLE A-2. Performance summary of EGR % control using NMPC

Engine speed [RPM] EGR RMSE Max Error

architecture [EGR%] [EGR%]

1000 HP-EGR 0.105 1.06

LP-EGR 0.13 0.66

DL-EGR 0.0489 0.6

2000 HP-EGR 0.13 1.53

LP-EGR 0.17 1.3

DL-EGR 0.07 0.061

3000 HP-EGR 0.16 2.07

LP-EGR 0.1 0.7

DL-EGR 0.2 2.34

Conclusion and future work

The proposed dual-loop EGR NMPC strategy is capable of coordinating the HP

and LP EGR mass flow rate to track the intake manifold pressure and intake manifold

EGR% references. Inclusion of turbocharger dynamics in the model facilitates optimal

balancing between the EGR loops whilst minimizing turbocharger lag. The NMPC

150
guarantees constraints and hence physically feasible values for the control inputs. It also

demonstrates the ability to consider the transport delay dynamics in the LP-EGR loop due

to the utilization of a multi-segment transport delay model. For the lower engine speeds,

DL-EGR control shows superior EGR% tracking performance. Inclusion of compressor

and turbine bypass valve, engine cylinder models in addition to further refinement of the

tuning parameters in the NMPC will be considered as future work.

151
 REFERENCES

[1] Lecointe, B. and Monnier, G., "Downsizing a Gasoline Engine Using Turbocharging

with Direct Injection," SAE Technical Paper 2003-01-0542, 2003,

https://doi.org/10.4271/2003-01-0542.

[2] Lake, T., Stokes, J., Murphy, R., Osborne, R. et al., "Turbocharging Concepts for

Downsized DI Gasoline Engines," SAE Technical Paper 2004-01-0036, 2004,

https://doi.org/10.4271/2004-01-0036.

[3] Zaccardi, J., Pagot, A., Vangraefschepe, F., Dognin, C. et al., "Optimal Design for a

Highly Downsized Gasoline Engine," SAE Technical Paper 2009-01-1794, 2009,

https://doi.org/10.4271/2009-01-1794.

[4] National Highway Transportation and Safety Administration- Corporate Average

Fuel Economy: https://www.nhtsa.gov/laws-regulations/corporate-average-fuel-

economy

[5] Shekaina, J., “Reduction Of Turbo Lag In Crdi Passenger Car And Afr Tuning Using

Neurofuzzy Controller”, PhD Thesis, Department Of Electronics And

Communication Engineering Dr. M.G.R. Educational And Research Institute

University.

152
[6] Jung, J., Park, C., and Bae, C., "Effects of High-Response TiAl Turbine Wheel on

Engine Performance under Transient Conditions," SAE Technical Paper 2015-01-

1881, 2015, https://doi.org/10.4271/2015-01-1881.

[7] Jung, J., Oh, H., and Bae, C., "Characteristics of Turbocharger with TiAl Turbine

Wheel in a Downsizing GDI Engine," SAE Int. J. Fuels Lubr. 6(3):579-588, 2013,

https://doi.org/10.4271/2013-01-2499.

[8] Steidten, T., Rauscher, M., Muenz, S.: Measures to improve the transient behavior of

turbochargers, ATK2008, p. 427-444, 2008

[9] Baines, N. C.: Fundamentals of Turbocharging, Concepts NREC, p. 94-95, 2005

[10] Park, S., Matsumoto, T., and Oda, N., "Numerical Analysis of Turbocharger

Response Delay Mechanism," SAE Technical Paper 2010-01-1226, 2010,

https://doi.org/10.4271/2010-01-1226.

[11] Siokos, K., Koli, R., “Assessment of simulation-based calibration for fuel

economy optimization”, International Journal of Powertrains

[12] Siokos, K., Koli, R., Prucka, R., Schwanke, J. et al., "Assessment of Cooled Low

Pressure EGR in a Turbocharged Direct Injection Gasoline Engine," SAE Int. J.

Engines 8(4):1535-1543, 2015, https://doi.org/10.4271/2015-01-1253.

[13] Ashok, B., Denis, S., Kumar, C., “A review on control system architecture of a SI

engine management system”, Annual reviews in Control, Vol. 41, 2016.

[14] Cholvin, R., “Turbocharger Controls”, SAE Summer Meeting Atlantic City, June,

1962

153
[15] Ementhal, et al, “Turbocharging Small Displacement Spark Ignition Engines for

Improved Fuel Economy”, SAE Technical Paper, 790311, 1979

[16] Karnik, A., Buckland, J., and Freudenberg, J., “Electronic Throttle and Wastegate

Control for Turbocharged Gasoline Engines”, 2005 American Control Conference,

June 8-10, 2005. Portland, OR, USA

[17] Lauber, K., Floquet, T., Colin, Y., “Nonlinear modelling and control approach for

a turbocharged SI engine”, IECON 2006 - 32nd Annual Conference on IEEE

Industrial Electronics

[18] Moulin, P., Chauvin, J., Youssef, B., “Modelling and control of the air system of

a turbocharged gasoline engine”, Proceedings of the 17th World Congress The

International Federation of Automatic Control Seoul, Korea, July 6-11, 2008

[19] Arnold, S., Groskreutz, M., Shahed, S., and Slupski, K., "Advanced Variable

Geometry Turbocharger for Diesel Engine Applications," SAE Technical Paper 2002-

01-0161, 2002, https://doi.org/10.4271/2002-01-0161.

[20] Hawley, J., Wallace, F., Cox, A., “Variable geometry turbocharging for lower

emissions and improved torque characteristics” Proceedings of the Institution of

Mechanical Engineers, Part D: Journal of Automobile Engineering

[21] Filipi, Z., Wang, Y., and Assanis, D., "Effect of Variable Geometry Turbine

(VGT) on Diesel Engine and Vehicle System Transient Response," SAE Technical

Paper 2001-01-1247, 2001, https://doi.org/10.4271/2001-01-1247.

154
[22] Brace, C., Cox, A., Hawley, J., Vaughan, N. et al., "Transient Investigation of

Two Variable Geometry Turbochargers for Passenger Vehicle Diesel Engines," SAE

Technical Paper 1999-01-1241, 1999, https://doi.org/10.4271/1999-01-1241.

[23] Petitjean, D., Bernardini, L., Middlemass, C., and Shahed, S., "Advanced

Gasoline Engine Turbocharging Technology for Fuel Economy Improvements," SAE

Technical Paper 2004-01-0988, 2004, https://doi.org/10.4271/2004-01-0988.

[24] Shimizu, K., Sato, W., Enomoto, H., and Yashiro, M., "Torque Control of a Small

Gasoline Engine with a Variable Nozzle Turbine Turbocharger," SAE Technical

Paper 2009-32-0169, 2009.

[25] Andersen, J., Karlsson, E., and Gawell, A., "Variable Turbine Geometry on SI

Engines," SAE Technical Paper 2006-01-0020, 2006, https://doi.org/10.4271/2006-

01-0020.

[26] Tang, H., Pennycott, A., Akehurst, S., Brace, C., “A review of the application of

variable geometry turbines to the downsized gasoline engine”, International Journal

of Engine Research 2015, Vol. 16(6) 810–825

[27] Martinez, R., Pesiridis, A., Yang, Y., “Overview of boosting options for future

downsized engines”, M. Sci. China Technol. Sci. (2011) 54: 318.

https://doi.org/10.1007/s11431-010-4272-1

[28] Katrasnik, T., Rodman, S., Trenc, F., Hribernik, A., Medica, V., “Improvement

of the Dynamic Characteristic of an Automotive Engine by a Turbocharger Assisted

by an Electric Motor”, Transactions of the ASME Journal of Engineering for Gas

Turbines and power, April 2003, Vol. 125

155
[29] Takaki, D., Tsuchida, H., Kobara, T., Akagi, M. et al., "Study of an EGR System

for Downsizing Turbocharged Gasoline Engine to Improve Fuel Economy," SAE

Technical Paper 2014-01-1199, 2014, https://doi.org/10.4271/2014-01-1199.

[30] Potteau, S., Lutz, P., Leroux, S., Moroz, S. et al., "Cooled EGR for a Turbo SI

Engine to Reduce Knocking and Fuel Consumption," SAE Technical Paper 2007-01-

3978, 2007, https://doi.org/10.4271/2007-01-3978.

[31] Kumano, K. and Yamaoka, S., "Analysis of Knocking Suppression Effect of

Cooled EGR in Turbo-Charged Gasoline Engine," SAE Technical Paper 2014-01-

1217, 2014, https://doi.org/10.4271/2014-01-1217.

[32] Wei, H., Zhu, T., Shu, G., “Gasoline engine exhaust gas recirculation – A

review”, Applied Energy Volume 99, November 2012,

[33] Alger, T., Gingrich, J., Roberts, C., “Cooled exhaust-gas recirculation for fuel

economy and emissions improvement in gasoline engines”, International Journal of

Engine ResearchVolume: 12 issue: 3,

[34] Cha, J., Kwon, J., Cho, Y. et al., “The effect of exhaust gas recirculation (EGR)

on combustion stability, engine performance and exhaust emissions in a gasoline

engine”, KSME International Journal (2001) 15: 1442.

https://doi.org/10.1007/BF03185686

[35] Grandin, B., Ångström, H., Stålhammar, P., and Olofsson, E., "Knock

Suppression in a Turbocharged SI Engine by Using Cooled EGR," SAE Technical

Paper 982476, 1998, https://doi.org/10.4271/982476.

156
[36] Kaiser, M., Krueger, U., Harris, R., and Cruff, L., "“Doing More with Less” - The

Fuel Economy Benefits of Cooled EGR on a Direct Injected Spark Ignited Boosted

Engine," SAE Technical Paper 2010-01-0589, 2010, https://doi.org/10.4271/2010-01-

0589.

[37] Kiwan, R., Stefanopoulou, A., Martz, J., Surnilla, G. et al., "Effects of Differential

Pressure Sensor Gauge-Lines and Measurement Accuracy on Low Pressure EGR

Estimation Error in SI Engines," SAE Technical Paper 2017-01-0531, 2017,

https://doi.org/10.4271/2017-01-0531.

[38] Kiwan, R., Stefanopoulou, A., Martz, J., “Effects of differential pressure

measurement characteristics on low pressure-EGR estimation error in Si-engines”,

2016, IFAC (International Federation of Automatic Control)

[39] Liu, F. and Pfeiffer, J., "Estimation Algorithms for Low Pressure Cooled EGR in

Spark-Ignition Engines," SAE Int. J. Engines 8(4):1652-1659, 2015,

https://doi.org/10.4271/2015-01-1620.

[40] Koli, R., Siokos, K., Prucka, R., Jade, S. et al., "A Control Algorithm for Low

Pressure - EGR Systems Using a Smith Predictor with Intake Oxygen Sensor

Feedback," SAE Technical Paper 2016-01-0612, 2016, https://doi.org/10.4271/2016-

01-0612.

[41] Siokos, K., “Low-Pressure EGR in Spark-Ignition Engines: Combustion Effects,

System Optimization, Transients & Estimation Algorithms”, PhD Thesis, Clemson

University, 2017.

157
[42] Guzzela, L. and Onder, C., “Introduction to Modeling and Control of Internal

Combustion Engine Systems” Springer publications

[43] Chung, J., Lee, S., “Measurement of Inertia of Turbocharger Rotor in a Passenger

Vehicle”, Transactions of the Korean Society of Automotive Engineers 2016

[44] Leufven, O., and Eriksson, L., “Surge and Choke Capable Compressor”, Model

2011 IFAC World Congress, Milano, Milano

[45] Leufven, O., and Eriksson, L., “A surge and choke capable compressor flow

model—Validation and extrapolation capability”, Control Engineering Practice, 2013

[46] Leufven, O., and Eriksson, L., “Measurement, analysis and modeling of

centrifugal compressor flow for low pressure ratios”, International Journal of Engine

Research, 2016

[47] Nelson, S., Filipi, Z., Assanis, D., “The Use of Neural Nets for Matching Fixed or

Variable Geometry Compressors with Diesel Engines”, Journal of Engineering for

Gas Turbines and Power, 1996

[48] Moraal, P. and Kolmanovsky, I., "Turbocharger Modeling for Automotive

Control Applications," SAE Technical Paper 1999-01-0908, 1999,

https://doi.org/10.4271/1999-01-0908.

[49] El Hadef, J., Colin, G., Chamaillard, Y., and Talon, V., "Physical-Based

Algorithms for Interpolation and Extrapolation of Turbocharger Data Maps," SAE

Int. J. Engines 5(2):363-378, 2012, https://doi.org/10.4271/2012-01-0434.

158
[50] El Hadef, J., Janas, P., Colin, G., Talon, V. et al., "Geometry-Based Compressor

Data-Maps Prediction," SAE Technical Paper 2013-01-0933, 2013,

https://doi.org/10.4271/2013-01-0933.

[51] Jensen, J., Kristensen, A., Sorenson, S., Houbak, N. et al., "Mean Value Modeling

of a Small Turbocharged Diesel Engine," SAE Technical Paper 910070, 1991,

https://doi.org/10.4271/910070.

[52] Sorenson, S., Hendricks, E., Magnusson, S., and Bertelsen, A., "Compact and

Accurate Turbocharger Modelling for Engine Control," SAE Technical Paper 2005-

01-1942, 2005, https://doi.org/10.4271/2005-01-1942.

[53] Eriksson, L., “Modeling and Control of Turbocharged SI and DI Engines”, Oil &

Gas Science and Technology – Rev. IFP, Vol. 62 (2007), No. 4, pp. 523-538

Copyright c 2007, Institut français du pétrole DOI: 10.2516/ogst:2007042

[54] Zhu, Q., “A Study Model Predictive Control for Spark Ignition Engine

Management and Testing”, PhD Thesis, Clemson University, 2015.

[55] Lee, J., “Model Predictive Control: Review of the Three Decades of

Development”, International Journal of Control, Automation, and Systems (2011)

9(3):415-424 DOI 10.1007/s12555-011-0300-6

[56] Rawlings, J., Muske, R., “The Stability of Constrained Receding Horizon

Control”, IEEE Transactions On Automatic Control, Vol. 38, No. 10, October 1993

[57] Colin, G., Chamaillard, Y., Charlet, A., Bloch, G. et al., "Linearized Neural

Predictive Control A Turbocharged SI Engine Application," SAE Technical Paper

2005-01-0046, 2005, https://doi.org/10.4271/2005-01-0046.

159
[58] Colin, G., Chamaillard, Y., Bloch, G., Corde, G., “Neural Control of Fast

Nonlinear Systems— Application to a Turbocharged SI Engine With VCT”, IEEE

Transactions On Neural Networks, Vol. 18, No. 4, July 2007

[59] Le Berr, F., Miche, M., Le Solliec, G., Lafossas, F. et al., "Modelling of a

Turbocharged SI Engine with Variable Camshaft Timing for Engine Control

Purposes," SAE Technical Paper 2006-01-3264, 2006, https://doi.org/10.4271/2006-

01-3264.

[60] Ericsson, G., Angstrom, H., and Westin, F., "Optimizing the Transient of an SI-

Engine Equipped with Variable Cam Timing and Variable Turbine," SAE Int. J.

Engines 3(1):903-915, 2010, https://doi.org/10.4271/2010-01-1233.

[61] “Exhaust Gas Recirculation Control for Large Diesel Engines - Achievable

Performance with SISO Design,” IFAC 9th Conference on Control Applications in

Marine Systems, ISSN – 14746670.

[62] Hegarty, K., Dickinson, P., Cieslar, D., and Collings, N., “Fast O2 Measurement

using Modified UEGO Sensors in the Intake and Exhaust of a Diesel Engine,” SAE

Technical Paper 2013-01-1051, 2013, doi:10.4271/2013-01-1051.

[63] Welling, O. and Collings, N., “UEGO Based Measurement of EGR Rate and

Residual Gas Fraction,” SAE Technical Paper 2011-01-1289, 2011,

doi:10.4271/2011-01-1289.

[64] Normey-Rico, J., “Control of Dead-time Processes,” Springer publications

[65] Heywood, J., “Internal Combustion Engine Fundamentals,” McGraw-Hill

publications

160
[66] Nishio, Y., Hasegawa, M., Tsutsumi, K., Goto, J. et al., "Model Based Control for

Dual EGR System with Intake Throttle in New Generation 1.6L Diesel Engine," SAE

Technical Paper 2013-24-0133, 2013, doi:10.4271/2013-24-0133.

[67] Fu, J. and Kurihara, N., "Intake Air Control of SI Engine Using Dead-Time

Compensation," SAE Technical Paper 2003-01-3267, 2003, doi:10.4271/2003-01-

3267.

[68] Siokos, K., Koli, R., Prucka, R., Schwanke, J., Jade, S.,“Physics-Based Exhaust

Manifold Pressure and Temperature Estimation for Low Pressure EGR Control in

Turbocharged Gasoline Engines,” – SAE Technical Paper 2016-01-0575, 2016.

[69] Zheng, M., Reader, G., Hawley, J., “Diesel engine exhaust gas recirculation––a

review on advanced and novel concepts”, Vol 45, Issue 6, Energy conversion and

Management, 2004

[70] N. Khaled, M. Cunningham, J. Pekar, A. Fuxman, and Ondrej Santin,

“Multivariable Control of Dual Loop EGR Diesel Engine with a Variable Geometry

Turbo”, SAE Technical Paper, No. 2014-01-1357, (2014).

[71] Wang, S., “Model Based Combustion Phasing Control for High Degree of

Freedom Spark-Ignition Engines”, PhD Thesis, Clemson University, 2015.

[72] Baratta, M., Spessa, E., and Mairone, P., “Numerical investigation of turbolag

reduction in hd cng engines by means of exhaust valve variable actuation and spark

timing control”, International Journal of Automotive Technology, Vol. 11, No. 3, pp.

289−306 (2010)

161
[73] Ferreau, H., Ortner, P., Langthaaler, P., del Re, L., Diehl, M., “Predictive control

of a real-world diesel engine using an extended online active set strategy”, Annual

Reviews in Control, vol. 31, no. 2, pp.293-301, 2007.

[74] Ortner, P., Langthaaler, P., Ortiz, J., del Re, L., “MPC for a diesel engine air path

using an explicit approach for constraint systems”, in Proc. Of 2006 IEEE International

Conference on Control Applications, pp. 2760-2765, Munich, Germany, Oct., 4-6,

2006.

[75] Garcı´a-Nieto, S., Martı´nez, M., Blasco, X., Sanchis, J., “Nonlinear predictive

control based on local model networks for air management in diesel engines” Control

Engineering Practice 16 (2008) 1399– 1413

[76] Emekli, Aksun, “Explicit MIMO Model Predictive Boost Pressure Control of a

Two-Stage Turbocharged Diesel Engine”, IEEE transactions on control systems

technology, vol. 25, no. 2, march 2017

[77] Wiese, A., Stefanopoulou, A., Karnik, A., Buckland, J., “Model Predictive

Control for Low Pressure Exhaust Gas Recirculation with scavenging” 2017

American Control Conference.

[78] Hu, Y., Chen, H., Ping, W., Chen, H., Ren, L., “Nonlinear model predictive

controller design based on learning model for turbocharged gasoline engine of

passenger vehicle”, Mechanical Systems and Signal Processing, Vol. 109, 2018.

[79] Zhu, Q., Onori, S., and Prucka, R., “Nonlinear economic model predictive control

for SI engines based on sequential quadratic programming”, in Proc. of 2016

American Control Conference, Jul., 2016, Boston, USA

162
[80] Zhu, Q., Onori, S., and Prucka, R., “Pattern recognition technique based active set

QP strategy applied to MPC for a driving cycle test”, in Proc. of 2015 American Control

Conference, pp. 4935~4940. Jul., 2015, Chicago, USA.

[81] Zhu, Q., Onori, S., and Prucka, R., “Nonlinear economic model predictive control

for SI engines based on sequential quadratic programming”, in Proc. of 2016 American

Control Conference, Jul., 2016, Boston, USA.

[82] Bemporad, A., Bernardini, D., Long, R., and Verdejo, J., "Model Predictive

Control of Turbocharged Gasoline Engines for Mass Production," SAE Technical

Paper 2018-01-0875, 2018, https://doi.org/10.4271/2018-01-0875.

[83] Zhu, Q., Koli, R., Feng, L., Onori, S., Prucka, R., “Nonlinear Model Predictive Air

Path Control for Turbocharged SI Engines with Low Pressure EGR and a Continuous

Surge Valve”, 2017 American Control Conference.

[84] Nicolau, G., Scattolini, R., Siviero, C., “Modelling the volumetric efficiency of ic

engines: parametric, non-parametric and neural techniques”, ControlEng. Practice,

Vol. 4, No. 10, pp. 1405-1415, 1996

[85] Wu, B., Filipi, Z., Assanis, D., Kramer, D. et al., "Using Artificial Neural Networks

for Representing the Air Flow Rate through a 2.4 Liter VVT Engine," SAE Technical

Paper 2004-01-3054, 2004, https://doi.org/10.4271/2004-01-3054.

[86] Matlab neural network toolbox https://www.mathworks.com/products/neural-

network.html

163
[87] Lee, T., Filipi, Z., “Nonlinear model predictive control of a dual-independent

variable valve timing engine with electronic throttle control”, Proceedings of the

Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering, 2011

[88] Vahidi, A., Stefanopoulou, A., “Current Management in a Hybrid Fuel Cell

Power System: A Model-Predictive Control Approach”, IEEE transactions on Control

Systems Technology, Vol.:14 Issue:6.

[89] Thomasson, A., Llamas, X., and Eriksson, L., "Turbo Speed Estimation Using

Fixed-Point Iteration," SAE Technical Paper 2017-01-0591, 2017,

https://doi.org/10.4271/2017-01-0591.

[90] Antoniewicz, K., Jasinski, M., Kazmierkowski, M., Malinowski, M., “Model

Predictive Control for Three-Level Four-Leg Flying Capacitor Converter Operating as

Shunt Active Power Filter”, IEEE transactions on industrial electronics, vol. 63, no. 8,

august 2016

[91] Hoffman, K., Heßeler, F., Abel, D., “Rapid Control Prototyping with Dymola and

Matlab for a Model Predictive Control for the air path of a boosted diesel engine”, E-

COSM - Rencontres Scientifiques de l'IFP - 2-4 Octobre 2006, Proceedings, pp. 33-40

Copyright ©2006, Institute for Automatic Control, RWTH Aachen, Germany

[92] Source of engine picture:

https://www.gmperformancemotor.com/parts/19328837.html

[93] X. Zhou, Y. Li, Y. Hu, and H. Chen, “Torque tracking control of turbocharged

gasoline engine using nonlinear MPC”, In 2015 European Control Conference, pp.

2958-2963, (2015).

164
[94] H. Borhan, G. Kothandaraman, and B. Pattel, “Air handling control of a diesel

engine with a complex dual-loop EGR and VGT air system using MPC”, In 2015

American Control Conference, pp. 4509-4516, (2015).

[95] B. Houska, H.J. Ferreau, and M. Diehl. ACADO Toolkit – An Open Source

Framework for Automatic Control and Dynamic Optimization. Optimal Control

Applications and Methods, 32(3):298–312, 2011.

165