Optimization of a Predictive Drive Strategy for a Plug-In Hybrid Vehicle Optimierung der vorausschauenden Antriebssteuerung bei einem Plug-In Hybrid

Bastian GINDROZ

Master of Science Thesis

at the Royal Institute of Technology, KTH Department of Vehicle Engineering Stockholm, Sweden

Dates 16 September 2013 - 14 March 2014 Firm AG Ingolstadt I/EA-24 Local supervisor Boris BLASINSKI Supervisor at KTH Daniel WANNER Examiner at KTH Lars DRUGGE Supervisor at ISAE-SUPAERO Joseph MORLIER

Abstract As the world’s natural resources are depleting while the environment constraints on society become om- nipresent, new technological alternatives for transports are sought. Over the past years, the regulations in terms of pollutants stimulated the industries’ innovation in almost all regions of the world, using directives such as the European emission standards have managed to limit the carbon dioxide emissions up to 82% between Euro 1 and Euro 4. Between research towards new fuels and the improvement of internal combustion engines (ICE), one solution stands out: combining an electrical machine with a traditional ICE to reduce the environmental impact and the fuel consumption. The most advanced level of hybridization is the Plug-In Hybrid Vehicle (PHEV), which usually has a battery capacity allowing it to drive all electric on short ranges and can be recharged at a power socket. This technological solution opens to a wide range of possibilities and improvements in terms of fuel consumption, as well as commercial opportunities. This master thesis, conducted at Audi AG in Ingolstadt () will focus on one development feature of Plug-In Hybrid Vehicles: the predictive drive strategy. Driving all electric everywhere would be the best solution in terms of fuel economy. However, a PHEV does not always allow it, especially on long rides. Mainly three drive modes are available: an electric mode that will use the ICE as seldom as possible, a hybrid mode that allows electrical experience at low speed and uses the ICE for higher needs in power, and a charge mode aiming at recharging the battery through the ICE. Some enhanced algorithms are also able to detect the different environments the vehicle is expected to drive through, and compute fuel and electrical consumption estimates for each modes available. Based on this information, an optimizer will determine the best drive strategy - i.e the best temporal combination of modes - in terms of fuel consumption, but also taking into account some external requirements, such as Zero Emission Zones that might come up along the route. The method used to compute the best drive strategy relies on four main stages. A first step is to identify the different driving environments constituting the route, based on long-range power predictions. This can be done using statistical comparison methods based on the assumption that one driving environment can be characterised by a range of values for a single parameter. The interval the parameter belongs to will determine the driving environment. The Kolmogorov-Smirnov test function is the second method used, which compares the cumulative distributions of two series of values and assesses their similarity. Once the itinerary is properly split into segments, it is possible to compute the expected fuel and electrical consumptions by taking the drivetrain efficiency chain, the regenerative braking and E-boost characteristics into account. For each section, a high- priority drive mode can also be given in order to allow a more flexible optimization in terms of coding and algorithmic. Mainly two prioritization options have been studied in this thesis. One pertains to the use a finite number of identified driving environments, to which a prioritization of the driving mode has been performed offline beforehand. The second option, the online priority, uses realtime computed values for a given segment and compares them in order to find the mode to be prioritized. Finally, two optimizers have been developed to compute an efficient drive strategy. The tree-based global optimizer considers the drive mode combinations for the whole route and selects the most efficient one, while a simplified optimizer would only improve a given basic strategy to reduce the fuel consumption. The results of the comparison methods show interesting perspectives for the Kolmogorov-Smirnov test func- tion, as it only tells two segments apart without trying to identify the environment they belong to, hence a faster implementation. The parametric method proved its efficiency when working with a dual-environment condition, although an extension to more than two environments would make the whole interval computation process much more complex. Then, the consumption computation processes have been precisely defined. Fi- nally, both optimizers gave interesting results. The global optimizer is very complex algorithm which requires considerable CPU resources. However, it provides absolute optima to the optimization problem that will be used to observe the trend, behaviour and accuracy of some drive strategies. The simplified optimizer is a suggestion of a faster algorithm that does not computes the best solution but rather a good one. If the quality of the resulting solution is very route-dependant, it can provide drive strategies very close to the best ones computes by the global optimizer.

i ii Thanks

I would like to thank all of those who contributed to this thesis and supported me during my stay in Ingolstadt. Warm thanks to my engineering schools in Stockholm (KTH) and in Toulouse (ISAE-SUPAERO) and their great teachers, for giving me the knowledge and passion without which I would never have been able to conduct this work. A special thank to my examiner Lars Drugge and my supervisor Daniel Wanner at KTH, whose support and help were simply flawless. I would like to thank all my colleagues from the department EA-242 at AUDI for always being friendly, caring and helpful. More particularly to my direct colleagues in the «Pr¨adiktive Betriebsstrategie» Team: Boris Blasinski, Markus Kaun, Florian M¨uhlfeldand Stefan Weinzierl. Finally, this master thesis would not even have existed without the devotion and strong support of my supervisor at Audi, Boris Blasinski. A big thank to you Boris, for our very interesting discussions and the role of advisor you perfectly played. Finally, I cannot forget to thank warmly my two families. My German host family, Angelika, Antonia, Joachim and Sugar for accommodating me in Ingolstadt during these six months and making me feel like home! And of course, thanks to my mother, father and brother for supporting me through all of my studies.

iii iv Contents

1 Introduction 1 1.1 Hybrid cars as an answer to a global trend...... 1 1.1.1 Environmental context and global trends...... 1 1.1.2 Regulations and standards...... 2 1.1.2.1 European regulations...... 2 1.1.2.2 Pollutant formation...... 4 1.1.3 Hybrid cars...... 5 1.1.3.1 Electricity generation...... 6 1.1.3.2 Hybrid technology...... 6 1.2 Hybrid technology at Audi...... 14 1.2.1 Audi AG...... 14 1.2.2 Hybrid technology...... 16 1.2.2.1 Electric Drivetrain Development in the R&D...... 16 1.2.2.2 Audi hybrid models and technologies...... 17

2 Towards the optimal PDS: introduction 21 2.1 Principle...... 21 2.2 State of the art...... 22 2.2.1 Research studies...... 22 2.2.2 In the industry...... 23 2.2.2.1 BMW: Predictive Power Management...... 23 2.2.2.2 BMW: Anticipatory Energy Management...... 24 2.2.2.3 Daimler: predictive drive strategy for E-trucks...... 24

3 Towards the optimal PDS: methods 25 3.1 Measurements and datasets...... 25 3.1.1 Test car: Hybrid...... 25 3.1.2 Measured parameters...... 25 3.2 Simulink model...... 27 3.2.1 Input: road profiles...... 27 3.2.2 Power calculation...... 29 3.2.3 Gearbox...... 30 3.2.4 Energy Management System...... 31 3.2.5 Consumption and emissions...... 31 3.3 Energy vs Power distribution...... 32 3.4 Segmentation...... 33 3.4.1 Kolmogorov-Smirnov test...... 34 3.4.2 Statistical parametric comparison...... 35 3.5 Fuel and electrical consumptions...... 36 3.6 Priorities...... 39

v CONTENTS CONTENTS

3.6.1 Offline prioritization...... 39 3.6.2 Online prioritization...... 41 3.7 Optimizer...... 41 3.7.1 Global Optimizer...... 42 3.7.2 Simplified Optimizer...... 44

4 Towards the optimal PDS: results 47 4.1 Segmentation...... 47 4.2 Fuel and electrical consumptions...... 54 4.2.1 Index i0 and i1 ...... 54 4.2.2 Conversion values...... 56 4.3 Optimal drive strategy...... 57 4.3.1 Global optimum...... 57 4.3.2 Enhanced basic strategy...... 61 4.4 Resulting fuel savings...... 63

5 Conclusion 65 5.1 Predictive drive strategy implementation...... 65 5.2 Future work...... 66

6 Bibliography and tables 68

7 Appendix 77 7.1 Complementary Information about Audi...... 77 7.1.1 The car fleet...... 77 7.2 Algorithm implementations...... 79 7.2.1 Global optimizer: C-code...... 79 7.2.1.1 Tree construction: TreeBuilding.h and TreeBuilding.c ...... 79 7.2.1.2 Tree analysis: TreeAnalysis.h and TreeAnalysis.c ...... 81 7.2.1.3 Main: main.c ...... 82 7.2.2 Global optimizer: ASCET...... 84

vi Chapter 1

Introduction

1.1 Hybrid cars as an answer to a global trend

1.1.1 Environmental context and global trends Over the past decades, mobility has become a central concept for the society. Being able to move independently, quickly and almost anywhere is today more than just a possibility. It is a necessity. Even if several solutions exist to transport people from one place to another, ground vehicles, such as trains, cars and buses, are today the most affordable and flexible option that is offered for short- and mid-range travels. However, public transportation allows less flexibility and freedom than individual transportation. Individualism is also slowly taking over social bounds, so that travelling alone or with a restricted group of acquaintances has become a common and logical luxury. The consequence of these attributes of a modern society is that private means of transportation are today more popular than ever. Cars represent an appropriate answer to these desires. According to recent figures [1], the world’s vehicle population exceeded one billion units already in 2010. Even though this number also includes trucks and buses, cars would represent 70% of this amount. The emerging countries are, to this extent, a quite interesting indicator of the world dynamic. A widely used example is of course China, with its 240 million vehicles1 in 2013. If this only represents some 86 vehicles per thousand inhabitants, which is way below USA’s 800 vehicles per 1000 persons [2], some projections show that the Chinese level will top today’s American one by 2030 [3]. Most of automakers aim at taking advantage of this trend, whose effects are already visible on their sales records. For instance, Audi sold 21% more vehicles in China in 2013 than in the previous year [4], breaking its all time record in this region of the world with almost 500 000 deliveries. A large majority of the vehicles on the roads is equipped with traditional internal combustion engines. The increasing petrol demand will eventually use up all of the existing oil resources. The only question is about when it will occur. Exact figures regarding the world remaining oil resources do not exist, although some estimates reckon that the reserves are expected to deplete within 40 to 120 years. Transportation is in the front line and accounts for more than 61% of the yearly petroleum consumption [5], 85% of which for land transportation. During the conversion of gasoline or diesel to mechanical energy through the combustion process, greenhouse gases (GHG), toxic chemicals and irritating particles are released. According to the European Commission, 12% of Europe’s carbon dioxide emissions stem from cars, which is somewhat problematic as CO2 has one of the biggest contributions to the greenhouse effect among all GHG. In comparison, road transport has the second largest share in Europe, a few percents behind power generation. The combustion from a chemical point of view will be dealt with while discussing the emission regulations and standards in subsection 1.1.2. Hybrid cars, which are vehicles powered by two types of energies, come often as an environmental solution. A description of the electrical hybrid technologies will follow in subsection 1.1.3, with a special focus on Audi and its hybrid developments in section 1.2.

1Statement of the Chinese Ministry of Public Security, 30/01/2013

1 CHAPTER 1. INTRODUCTION 1.1. HYBRID CARS AS AN ANSWER TO A GLOBAL TREND

Chapter2 will first present the principle of a predictive drive strategy in section 2.1 and then a state of the art in the research publications and the industry in section 2.2. The methods used to compute a predictive drive strategy will be presented in3, followed by the results in chapter4. Finally, the conclusion chapter5 will wrap up the thesis while suggesting directions for future works in the area of predictive drive strategies.

1.1.2 Regulations and standards Considering the extent of these environmental and economic challenges, regulations have been introduced. However, such changes in the industry and the people’s mindset are not instantaneous, and always present an economic risk for the automakers. If the innovation will be driven by the research and development departments of each car company, nothing will ever happen without a strong regulatory context ruling and monitoring this evolution. Therefore, regulations and standards aim at controlling, stabilizing and possibly decreasing the scale of these challenges.

1.1.2.1 European regulations As far as emissions are concerned, two tools are usually used by the relevant institutions: • Regulation targets, applied to automakers’ strategies. They set a goal to reach and the time extent allowed for fulfilling the target. • Regulation standards focus more specifically on the emission performance of each vehicle. They have a defined date as of which they take effect and a set of emission levels that each car has to follow. Even if an international emission regulation rule does not exist today, almost all vehicles of the developed countries obey to some emission regulation sets. European emission standards are among the most widespread regulations in the world. Elaborated and set by the European Union by means of directives since 1991 [6], the Euro regulation standards are not limited to the old continent anymore. Many countries model on this set of directives to implement their own, especially developing countries such as China [7] and India [8].

Table 1.1: European emission standards for a passenger car (g/km), Diesel.

Tier Date CO THC NMHC NOx HC+NOx PM Euro 1 July 1992 2.72 - - - 0.97 0.14 Euro 2 January 1996 1.0 - - - 0.7 0.08 Euro 3 January 2000 0.64 - - 0.50 0.56 0.05 Euro 4 January 2005 0.50 - - 0.25 0.30 0.025 Euro 5 September 2009 0.50 - - 0.180 0.230 0.005 Euro 6 September 2014 0.50 - - 0.080 0.170 0.005

Table 1.2: European emission standards for a passenger car (g/km), Gasoline.

Tier Date CO THC NMHC NOx HC+NOx PM Euro 1 July 1992 2.72 - - - 0.97 - Euro 2 January 1996 2.2 - - - 0.5 - Euro 3 January 2000 2.3 0.20 - 0.15 - - Euro 4 January 2005 1.0 0.10 - 0.08 - - Euro 5 September 2009 1.0 0.10 0.068 0.060 - 0.005 Euro 6 September 2014 1.0 0.10 0.068 0.060 - 0.005

2 CHAPTER 1. INTRODUCTION 1.1. HYBRID CARS AS AN ANSWER TO A GLOBAL TREND

As shown on Tables 1.1 and 1.2, the regulations have become stricter since the introduction of Euro 1 in 1992, stressing the growing necessity to reduce the ecological footprint of vehicles. The pace of emission reductions and worldwide long-term plans enabled the European Union to set regulation targets. These directives aim at setting a framework for the automotive industry and the coming passenger car fleets. Regulation 443/2009 [9] has for instance defined a two-stage process to reduce GHG emissions of 20% by 2020. Each car manufacturer will get an annual specific emission target based on its average fleet weight [10]:

Specific emission of CO2 = 130 + a × (M − M0) (1.1) where: M is the average mass of the manufacturer’s fleet (in kg) M0 is the reference mass (1372.0 kg in 2012) a is 0.0457 For instance, Audi AG’s average mass was 1579.0 kg in 2012, which led to a 2012 target of 139 g/km CO2. However, individually, the car manufacturer had already comfortably met this requirement with an average CO2 emission of 122 g/km [10]. The average of the specific CO2 emissions over all manufacturers is nothing but the EU 2015 target of 130 g CO2/km. Until 2015, only a certain percentage of each manufacturer’s fleet will be taken into account for the average (65% in 2012, 75% in 2013, 80% in 2014 and 100% in 2015-2016). During this period of time, the phase-in will only concern the best-performing registered cars. Super-credits are also given to vehicles emitting less than 50 g CO2/km, giving a bigger weight to electric or very efficient hybrid vehicles in the fleet average. Even if this super-credit system has been criticized for misrepresenting the real carbon performance of car manufacturers [11], they constitute one more tool in favour of car hybridization. Audi AG, with an individual target of 139 g/CO2 is one of the 9 manufacturers that had already met their 2015 specific CO2 emission target in 2012. Audi AG is currently part of a joint pool with the VW group. The pool system allows manufacturers to form an alliance that will be considered as one entity with one specific CO2 emission target. Table 1.3 shows the different performances within the VW pool. By doing that, the advance of some car manufacturers regarding their target might be used to compensate other high-carbon fleets within the pool. This is often used between regular and sport brands, like in the BMW Group pool between Bayerische Motoren Werke AG and BMW M GmbH. Table 1.3: Distance to target for the VW pool in 2012 [10].

Average emissions [g Target [g Distance to target [g Manufacturer CO2/km] CO2/km] CO2/km] Audi AG 122 139 -17 Audi Hungaria Motor KFT 137 134 3.7 Bentley Motors Ltd 310 181 129 Bugatti Automobiles S.A.S 539 157 382 Automobili S.p.A 344 144 199 Dr. Ing. h.c. F. Porsche AG 188 153 35 Quattro GmbH 219 147 72 Seat SA 115 127 -12 Skoda Auto AS 120 127 -6.6 Volkswagen AG 119 131 -12 VW Group PC 120 132 -12

The EU target for 2020 is 95 g CO2/km (already mentioned in the Regulation 443/2009), with a weight slope (see formula 1.1) of a=0.0333 that should be confirmed in an amendment to be issued soon. Based on their previous CO2 emission reductions, some car manufacturers are already on good track to reach this target, such as Toyota with an overall progress required in the fleet CO2 emissions of 3.4% by 2020 to be compared with the 4.6% that the Japanese firm has achieved over the past 5 years.

3 CHAPTER 1. INTRODUCTION 1.1. HYBRID CARS AS AN ANSWER TO A GLOBAL TREND

Some local and national initiatives exist as well to further reinforce the European emission policy. Several governments financially support the development of low emission vehicles, either by funding R&D programs such as in the UK, or by offering a range of purchase rebates depending on the car’s CO2 emission level, like in France and Spain. Finally, some countries also grant EV and HEV vehicles very attractive advantages such as road tax and toll exemption and free parking in the cities [12]. The combination of regulation standards and targets creates a monitoring context. Emissions restrictions aim at controlling both current car environmental performances and the automotive industry’s strategy in the medium term. The requirements of such an ambitious regulation tactic would not have been met without tremendous improvements in the engine efficiency, as well as exhaust treatment systems.

1.1.2.2 Pollutant formation

Internal combustion engine functioning is based on the transformation of chemical energy into mechanical work. The chemical energy is provided by the combustion of fuel with air, during which reagents are consumed and products appear. As the combustion takes place in the combustion chamber, these newly created components will be evacuated through the exhaust and ultimately sent out in the atmosphere. Classically, the complete combustion of gasoline in the air could be written as follows:

C8H18 + 12.5 (O2 + 3.773N2) → 8CO2 + 9H2O + 47.16N2 (1.2)

However, combustions in traditional engines are far to be complete and perfect. The type and proportion of exhaust gases depend on the combustion reaction parameters: temperature, type of fuel, reagent proportions. For fuels such as gasoline and diesel, the main emissions are unburned hydrocarbons (HC), nitrogen oxides (NOx), carbon monoxide (CO) and soot [13]. In addition to these, nitrous oxide (N2O), methane (CH4), ammoniac (NH3) and urea will also be found. Biodiesel fuels usually leave less unburned hydrocarbons because the fuel already contains oxygen. However, the NOx proportions increase with that type of fuel. The main exhaust gases and their effects on health and nature are described below:

• Nitrogen oxides: they are produced in high temperature conditions, provided that there is nitrogen and enough oxygen to oxidise it [14]. Because NOx react with moisture and sunlight, then can cause respiratory problems when inhaled. NOx emissions also lead to a mechanism which destroys ozone in the stratospheric layers. • Carbon monoxide: CO is produced in big quantities during the combustion of rich fuel/air mixtures, because of a lack of oxygen to form CO2. The risks for health are significant, especially since this gas has no specific smell or color. When inhaled, CO reduces the ability of the body to convey oxygen, as hemoglobins are affected. On a larger scale, carbon monoxide is a short-life atmospheric pollutant which contributes to form smog in urban areas. • Hydrocarbons: HC come from completely unburned fuel, for one third [15], and partly burned fuel. Some of them end up in the exhaust gases due leaks and quenching phenomena. Such chemical components are toxic and carcinogenic to humans, and may lead to lungs and heart diseases. They also help in the formation of low level ozone, which both generates smog and an enhanced green house effect. • Particle matter: particle matter and soot are typical of Diesel combustions. Because they are bigger than the other pollutant existing in the exhaust gases, they can cause lung irritations and lead to cancers. Their effect on nature is also devastating, as such particles indirectly block the photosynthesis process [13]. • Carbon dioxide: CO2 is one common and natural gas. It comes as a normal combustion product (together with water). Carbon dioxide is also one major green house gas and leads to ocean acidification. However, CO2 does not represent a threat for nature and human in normal concentrations. • Sulfur oxides: Fuels generally contain sulfur compounds, which are oxidized and released in the atmo- sphere after the combustion. Sulfur oxides are major contributors to rain and ocean acidification.

4 CHAPTER 1. INTRODUCTION 1.1. HYBRID CARS AS AN ANSWER TO A GLOBAL TREND

• Nitrous oxide: adding N2O in the combustion chamber increases the level of oxygen, allowing to burn more fuel for a more powerful combustion. Nitrous is mostly used in racing cars. But it plays also a potent role in the greenhouse effect and the air pollution. Nitrous oxide is also known as laughing gas. • Methane: CH44 emissions come from incomplete fuel combustions [16]. Methane has a significant impact on the environment, as a significant greenhouse gas. • Ammoniac: In a SCR (Selective Catalytic Reduction) system, the thermolysis and hydrolysis of urea produces NH3 [17]. A failing in the SCR could cause a release of ammoniac through the exhaust. This chemical compound reacts with organic matter, such as human skin, which makes it dangerous to manip- ulate without accurate protection. • Urea: Injected shortly upstream the SCR catalyst, urea’s aim is to transform the NOx into ammoniac. The NH3 will then be involved in a reaction (intermediate, fast or slow) that will eventually consume it to produce water and dinitrogen. Urea is not particularly dangerous for humans, as the body produces this molecule. Tremendous technological improvements have stemmed from the need of a cleaner exhaust, confirming the stimulating role of environmental standards for the car industry.

Gasoline GDI, Gasoline Direct Injection, is based upon a fuel injection directly inside the combustion chamber. The idea is to obtain a better control of the combustion by modifying parameters shortly before, during and after the combustion. Significant advantages of this technology are the regulation of the combustion temperature and the possibility to stratify the charge [15]. The mix could then be around lambda 1 near the spark, meaning that the air and fuel are close to the stoichiometric proportions, and lean everywhere else. Besides, as the fuel is directly injected inside the cylinder, it evaporates there too. Since an evaporation consumes energy (heat in this case), the combustion will be colder. This reduces the risk of knocking and the production of NOx as well. A 3-way catalytic converter is a post-cylinder system that oxidizes HC and CO and reduces NOx. To work properly, the combustion needs to be close to lambda 1. Basically, the catalyst stores oxygen and uses it to reduce the hydrocarbons and the carbon monoxide.

Diesel Some techniques used to combat the NOx emissions for Diesel engines are [18]: • injection delay, leading to a later combustion during the expansion, hence lowering the temperature and the amount of NOx produced. • use of a pre-chamber in which the combustion starts and where the the conditions are stoichiometric due to the low amount of air available. However, smokes produced during the start of combustion will need to be eliminated later on. • intake air cooling, usually through an intercooler, which will lower the combustion maximal temperature and thus reduce the NOx emissions. • Exhaust Gas Recirculation (EGR), which consist of mixing exhaust gases to the intake air. This dilution will reduce the oxygen concentration and lower the combustion temperatures. • a selective catalytic reduction in which Urea is injected to reduce efficiently NOx into N2 [19]. Moreover, particulate matter and soot concentrations can be significantly reduced by using particle filters. These filters will either heat and inject fuel to burn the soot externally, or use internal disposable coating to catch soot particles.

1.1.3 Hybrid cars If the previously mentioned exhaust controls strategies have had a significant effect on the environmental impact of a car, treating the problem at its source is theoretically much more interesting. This thought led to the investigation of new alternative fuels as well as other sources of energy. Biofuels, alcohol fuels, hydrogen, compressed air have better environmental performances in terms of emissions. By replacing the tank with a

5 CHAPTER 1. INTRODUCTION 1.1. HYBRID CARS AS AN ANSWER TO A GLOBAL TREND

battery, a no-emission solution for transports would be offered through the use of an electrical machine, though not totally a clean product when considering the footprint of electricity generation. A compromise solution consists of combining an ICE and a electric drivetrain to reduce the fuel consumption as well as meeting the emission standards dictated on European and local scales.

1.1.3.1 Electricity generation The most widespread and industrially used electricity generation solutions are nuclear powerplants, fossil fuel power stations, hydropower, wind and solar energy. Fuel per kWh is cheaper than electricity, approximately half its price in Germany. However, with the bigger efficiency of electrical systems, a larger part of the incoming energy (electricity) is converted to outgoing energy (mechanical energy for engines). The energy savings are such that they compensate a higher price at the source, which in hybrid and electric cars enables a lower cost per km for electric powertrains than with traditional internal combustion engines.

1.1.3.2 Hybrid technology The term hybrid refers to a configuration where two power sources coexist within the vehicle in new propulsion systems technology, : fuel stored in a tank and electricity stored in a battery. The different degrees of hybridization and their features are presented in Table 1.4.

Table 1.4: Hybridization degrees [20].

Plug-In Micro hybrid Mild hybrid Full Hybrid Hybrid Start & Stop X X X X E-boost X X X Regenerative X X X braking Engine load X X point shifting Electric only X X Electric only X > 5 km Charging on X the grid

Hybrid vehicles have several combined features that enable comfort and energy savings. On the following pictures, the energetic distribution of these different features are shown. The ICE is located at the front of the car (left-hand side). The battery is represented by a pink rectangle at the rear (right-hand side). The energy coming from the battery is pink, the energy coming from the ICE is cyan and the energy going to the battery (during regenerative braking for instance) is yellow. The arrows indicate the direction of energy transfer.

Start & Stop This system equips a large part of recent vehicles. It turns off the engine when the car stops - at traffic lights and tolls for instance - and starts it again when the driver pushes the gas or clutch pedal. Of course, this general rule has plenty of technical exceptions preventing the engine from being turned off in critical situations, such as a high power requirement on the air conditioning system. To further improve this functionality, a driving situation recognition has been implemented at Audi. With this technology, the car automatically detects the type of driving environment - traffic jam for instance - and sets the Start & Stop parameters accordingly [21]. Finally, some hybrid cars also coast by shutting the engine off during some deceleration phases.

6 CHAPTER 1. INTRODUCTION 1.1. HYBRID CARS AS AN ANSWER TO A GLOBAL TREND

E-boost In case of a need of extra power at the wheels - overtaking for instance -, both the electrical machine and the ICE can be activated at the same time to meet the performance requirements. Figure 1.1 shows the energetic sources during an E-boost.

Figure 1.1: Schematic energetic diagram of the E-boost [22].

Regenerative braking An electric motor can also act as a power generator when the rotor spins at a speed greater than the syn- chronous speed (negative slip). This property is due to the reversible nature of the electro-magnetic phenomena involved. During a deceleration, an active braking control will decide the proportion of mechanical brakes to be used. The complementary part will be taken care of by the electric motor, hence generating energy and sending it back to the battery, as shown in Figure 1.2. However, this feature is mainly active for braking producing up to 40 kW approximately.

Figure 1.2: Schematic energetic diagram of the regenerative braking [22].

Engine load point shifting Being able to use the ICE at its most efficient load points is not always possible for traditional cars. To optimize the load point, the engine speed would for instance be slightly increased, maybe leading to a higher fuel consumption, but with a very interesting gain in power output. In hybrid vehicles, this extra power could be used to charge the battery at low cost. By analysing the engine efficiency map, the ECU (Engine Control Unit) shifts the load point to a more efficient area. Figure 1.3 shows the basic principle of a load point adjustment. On this torque vs engine speed map, the level curves represent iso-efficiency areas. The goal of a load point optimization is to move the load point to a local maximal efficiency zone. Most of the time, it implies an increase in the torque, such as shown by the grey arrow. The extra torque produced can then be stored as electric energy in the battery.

7 CHAPTER 1. INTRODUCTION 1.1. HYBRID CARS AS AN ANSWER TO A GLOBAL TREND

Figure 1.3: Engine map and effect of a load point increase/decrease [23].

The resulting consequences in terms of performance are then taken on by the electric drivetrain. If the load point has been lowered to a zone with a smaller torque than required, then the complementary energy will be taken out of the battery to compensate (see Figure 1.4(a)). If, on the other hand, the local optimum in terms of specific consumption is in a region with a higher torque, the extra output energy will be stored into the battery (see Figure 1.4(b)).

(a)

(b)

Figure 1.4: Schematic energetic diagram of a lower load point (a) and a higher load point [22].

8 CHAPTER 1. INTRODUCTION 1.1. HYBRID CARS AS AN ANSWER TO A GLOBAL TREND

Electric only The «electric only» feature is exclusively available on Full-hybrid cars (including Plug-In Hybrids) and allows the car to run on the electric machine without any contribution of the ICE (see Figure 1.5). The normal driving mode involving the ICE only is still available (Figure 1.6).

Figure 1.5: Schematic energetic diagram of the full electric mode [22].

Figure 1.6: Schematic energetic diagram of the normal drive [22].

As this work deals with Plug-In Hybrid technology, most of the oncoming studies will focus on this technology rather than other degrees of hybridization. In a Full Hybrid or Plug-In Hybrid vehicle, the hybrid technology can be implemented through several architectures. A basic classification gives three main types of drivetrains: • series hybrid, such as the e-tron • parallel hybrid, such as the Hybrid • power-split, such as the Toyota Prius Note that other architectures exist, like the axle-split technology used on the BMW i8 which is powered by the electric machine on the front axle and by an ICE on the rear axle. All architectures have most of the components in common: • Batteries: to store electrical energy. There are usually two batteries: one is a high voltage battery aiming at supplying high voltage current for the hybrid functions, such as E-driving and regenerative braking. The second battery is typically a 12-Volt Battery that supplies the onboard electrical network and its features (air conditioning, radio, displays, . . . ). • Converter: usually a DC/AC converter, adapts the voltage of the current coming out of the high-voltage battery to the electrical machine, as well as the current from the machine to the high-voltage battery. • Electric motor: 3-phase AC induction motor, 145 V for the Audi Q5 Hybrid [24]. This type of motor is made of two main parts: a rotor, usually a cylindrical metal cage with conducting materials in its surface, spins under the effect of a rotating electro-magnetic field generated by the stator.

9 CHAPTER 1. INTRODUCTION 1.1. HYBRID CARS AS AN ANSWER TO A GLOBAL TREND

Series hybrid The series hybrid architecture specificity is that the electric machine is the only engine that is directly coupled to the wheels. The role of the ICE is to supply the electric drivetrain with power, to the battery or directly to the electric motor, as shown on Figure 1.7. An energy storing system, which is optional for this architecture enables regenerative braking. This architecture is also used for electric vehicles equipped with a range extender (EREV2) . A range extender is an ICE that aims at increasing the range of an electric car. In 2010 Audi presented its A1 e-tron at the Geneva International Motorshow. The A1 e-tron is a range-extender equipped version of the A1. To support an electric machine of 45 kW, a Wankel engine producing 15 kW has been installed. Although its performance is limited, the range extender can bring 150 extra kilometres of range [25]. The main advantage of a series hybrid architecture is that the ICE can be run at its best load point most of the time [20]. As it does not directly drive the wheels, the engine speed and torque can be chosen freely depending on the electrical needs. Any extra energy generated can be stored in the battery, whereas a lack of power would be compensated by the very same battery. Moreover, as the electric machine benefits from a direct access to the wheels, the gearbox is drastically simplified while regenerative braking presents an enhanced efficiency. However, these vehicles have a lower overall fuel efficiency due to the energy conversion losses. The complex architecture implies a heavy drivetrain with a sustained demand on both engines to allow a reasonable speed range.

Figure 1.7: Schematic diagram of a series hybrid drivetrain [23].

2Extended Range Electric Vehicle.

10 CHAPTER 1. INTRODUCTION 1.1. HYBRID CARS AS AN ANSWER TO A GLOBAL TREND

Parallel hybrid The parallel hybrid is one of the most common architectures among the hybrid vehicles. It differs from the series architecture by the fact that both the ICE and electric machine are able to directly power the wheels (see Figure 1.8). Differential gears are here to ensure the torque transmission to the wheels as well as the coordination between electric machine and ICE. The Audi Q5 Hybrid is an example of a parallel hybrid. That SUV is equipped with a 180 kW four-cylinder engine, assisted by a 40 kW 3-phase electric machine supplied by the 1.3 kWh battery [24]. Another parallel architecture type is the «through the road» layout. In this case, the ICE powers one axle, usually the front one, whereas the electric machine powers the other one. This system is used on the Volvo V60 Plug-In Hybrid where the rear axle is electrified and the front one powered with a turbo diesel injection engine. Although no mechanical connection exists between the electric machine and the ICE, the architecture is still a parallel hybrid with all related features. Parallel hybrids have the advantages offering a better efficiency on long distance driving as well as a consid- erable flexibility regarding the equipment and operation. The electric machine does not need to be as big and powerful as for a series hybrid, as it only assists the ICE. On the other hand, because the ICE cannot be run at optimal load point at all time, parallel hybrids are less efficient in city driving.

Figure 1.8: Schematic diagram of a parallel hybrid drivetrain [23].

11 CHAPTER 1. INTRODUCTION 1.1. HYBRID CARS AS AN ANSWER TO A GLOBAL TREND

Power-split This solution shown on Figure 1.9 is an intermediary architecture between series and parallel. The mechanical coupling between the ICE and the electric motor through differential and planetary gears allows different types of power paths - either mechanical or electric. As a result, a series hybrid configuration can be used at low speed, by simply blocking the direct mechanical link existing between the internal combustion engine and the wheels. During cruising, a parallel layout will be used for maximum efficiency. The high flexibility and adaptability of this solution was one of the elements of the success of the Toyota Prius. However, the level of complexity of this technology is high, making it rather expensive to produce and buy [20].

Figure 1.9: Schematic diagram of a power-split hybrid drivetrain [23].

12 CHAPTER 1. INTRODUCTION 1.1. HYBRID CARS AS AN ANSWER TO A GLOBAL TREND

Battery and charging options All of these technologies rely on batteries, more specifically on high-voltage batteries. Their aim is to store energy to make it available for the electric machine when needed. Figure 1.10 shows the difference in energy density between the different fuels and main battery types. It is interesting to notice that batteries cannot compete with fuels in terms of energy per weight.

Figure 1.10: Energy density of some fuels and batteries [12].

A choice based on the application has to be made when selecting the battery type. An HEV vehicle will use the electric drivetrain mainly as a power support unit for E-boost for instance. The battery does not aim at providing a lot of energy as the all-electric mode of HEVs is quite limited. Consequently, a high power density battery will be preferred, with a capacity of approximately 1-2 kWh. NiMH batteries are widely used for this application. On the other hand, Plug-In Hybrids and EV cars are designed to provide a much larger electric experience than conventional hybrids. This means that the electric resources and features are used for a longer time and thus that the battery should be able to store a considerable amount of energy: For these applications, high energy density lithium batteries will be chosen, providing a capacity up to 85 kWh3. If 100% of a battery capacity can theoretically be used, at least for new batteries, the exploited range of State Of Charge (SOC) is narrower. The Depth of Discharge (DOD) plays a major role in the battery life. As no memory effect affects the Lithium-Ion batteries, it is absolutely acceptable to partially discharge it before charging it again. Usually, a lower stress level on the battery components - i.e a lower DOD - will increase the number of possible discharge cycles. Due to ageing of the cells, an old battery cannot be charged at 100% of its initial capacity. Moreover, as the battery capacity is not always absolutely exactly known, it is safer to define SOC boundaries that prevent to reach extreme states of charge. Finally, a Lithium-Ion battery is permanently controlled by a Battery Monitoring System (BMS) which would put the battery into sleep mode if the SOC become too low. For all of these reasons, energy management programs and systems always work with a restricted SOC range, between 20% and 80% for instance. For PHEVs and EVs the battery charge is a challenging topic. To compensate the limited range of these vehicles, the battery recharge should be as quick as possible. However, in some cases it can last up to 20 hours. As the market of battery systems is flourishing, the technological development is evolving at a fast pace. Several recharging technologies exist today and can basically be classified into two types:

• Plug-in charging • Wireless Power Charging (WPT)

Plug-In charging involves a power cable that connects the vehicle inlet to the charging facility. This option is today the most widely used for both public and domestic purposes. The EV or PHEV can then be charged while shopping at the mall or during the night at home. The IHS Inc.4 report about batteries [12] mentions three levels of plug-in charging technologies planned in the US:

3The Tesla Model S high capacity battery 4Formerly Information Handling Services Incorporated, global market and economic information provider.

13 CHAPTER 1. INTRODUCTION 1.2. HYBRID TECHNOLOGY AT AUDI

• 120V-15A AC source. Common for household purposes, the drawback of this technology is the duration of the recharging process. In the example of an EV such as the Nissan Leaf, it can take up to 19 hours, which brings scheduling issues. • 240V-30A source. This option halves the charging time and can be integrated directly at home, as some domestic devices already work on 240V. The technology is likely to be the most widespread one for EV and PHEV cars recharging in a very near future. • 480V three-phase DC. Industrial technology offering a high speed recharging. However, beyond the costs, it also needs specific plugs and charging facilities. The Nissan Leaf can be charged up to 80% in 26 minutes.

In order to avoid as much as possible compatibility issues related to plugs as well as limiting the onboard electronics dedicated to charging, a wireless charging method has been developed recently. WPT is based on inductive coupling: a coil located under the car carries an unsteady current and thus generates an electromagnetic field. The second coil installed in the car is close enough to the first one to be influenced by the generated electromagnetic field. Following the Neumann induction principle, this field induces a current in the second coil and consequently charges the battery. This technology reduces the issues of compatibility and electronics mentioned above, but raises new matters, especially regarding alignment with the underground coil and efficiency. Volvo is leading a project called Continuous Electric Drive which is based on the plugless charging of a C30 (see Figure 1.11).

Figure 1.11: Volvo’s Continuous Electric Drive system, Volvocars news webpage.

The battery capacity and the charging technologies will decide of the viability of PHEVs and EVs in the future. They are driving parameters of the electromobility development which keep being improved and chal- lenged. In parallel to that, a small revolution regarding the onboard power-net is also taking place. Most of today’s cars are supplied with 12V. However, more and more modern integrated devices demand a significant power supply, such as the CPT SpeedStart Starter generator [26] whose energy recovering and generation features are considerably boosted when powered at 48V. The technical and economical issue has triggered some big OEMs such as the VW Group, BMW and Daimler to start integrating a 48V subnet to their onboard systems, connected to the conventional 12V net via a two-way DC/DC converter.

1.2 Hybrid technology at Audi 1.2.1 Audi AG History August was born in western Prussia in 1868. Following a degree in engineering and some experience in shipbuilding, Horch decided to found his own business A. Horch & Co in at the age of 31. Two years later the first Horch automobile, a two-cylinder engine car, was produced. After moving his firm out of Cologne,

14 CHAPTER 1. INTRODUCTION 1.2. HYBRID TECHNOLOGY AT AUDI

Horch is forced to rename it following a court decision: Audi Automobilwerke GmbH is officially registered on the 25 April 1910, Audi being a latin transformation of Horch. In 1932, Audi, DKW, Horch and gather and form the four ring company AG. AUTO UNION GmbH is founded is Ingolstadt in 1949, becomes a Volkswagen owned subsidiary in 1966. NSU Motorenwerke AG joins the alliance three years later. The company is finally renamed AUDI AG, its current name, in 1985, and is today one of the leading international suppliers of premium automobiles.

Management and strategy

Audi is part of the VW Group, together with 11 other vehicle manufacturers, as shown on Figure 1.12. The firm is coordinated by a board of management of 6 directors and chaired by .

Figure 1.12: Chart of the VW Group and its subsidiaries.

Audi’s vision 2020 is to be the premium brand. The mission plan is oriented around 4 different axis:

• Living responsibly, by making the difference through social involvement, green and resource-efficient value chain and economic responsibility. • Defining innovation, via technological programs such as the Audi e-tron, Audi ultra5 and Audi design. • Creating experiences with innovative sales formats, by presenting the brand in an authentic and likeable way and providing personal customer relationships • Shaping Audi towards flexibility in high performance processes and structures, global presence and com- petence and agility for growth and new business areas

The Goals of Audi pertain to acquiring a superior financial strength, maintaining a continuous growth, becoming a global image leader as well as an attractive employer worldwide.

Key figures and facts

In 2013, Audi employed 52 563 persons [27], 37 400 of which working in Ingolstadt. The operating income of Audi amounted 5 billion euro in 2013 [28], with an increasing turnover of 49.9 billion euro that will contribute to the large 22 billion euro investment planned by 2018. In 2013, the Audi fleet was constituted 12 models giving more than 50 variants (Figure 7.1 in the appendix, section 7.1). Audi produced 1 605 926 units and sold 1 575 480 cars in 2013. 31% of these cars were sold in China, 35% in Europe and 10% in the USA. Figure 1.13 shows the model distribution in the 2013 sales.

5Low Weight Car Body Project

15 CHAPTER 1. INTRODUCTION 1.2. HYBRID TECHNOLOGY AT AUDI

Figure 1.13: Shares of Audi models in the 2013 sales [27].

1.2.2 Hybrid technology 1.2.2.1 Electric Drivetrain Development in the R&D The Research & Development Unit6 is located in Ingolstadt, Germany, on the main Audi facility where more than 37 000 people work. Among all the different departments gathered on this site, one, called EA7, deals with Powertrain Devel- opment and is chaired by Dr. Hirsch. Figure 1.14 gives an overview of the organization of the Entwicklung Aggregate Department. One of its sub-departments, the EA-2, is responsible for the development of electric drivetrains. EA-2 is lead by Mr Enzinger and also chaperons 4 groups, including the EA-24 in charge of the functions and applications for electric drivetrains. More specifically, a team of a twelve engineers works on the drive strategies under the name of EA-242. This thesis has been conducted within EA-242, which carries out projects pertaining to the Start & Stop functions, as well as predictive drive strategies.

Figure 1.14: Organization chart of the Powertrain Development Department at Audi, focus on the group EA-242 where the thesis has been conducted.

6Commonly referred to as TE, which stands for Technische Entwicklung 7Entwicklung Aggregate

16 CHAPTER 1. INTRODUCTION 1.2. HYBRID TECHNOLOGY AT AUDI

1.2.2.2 Audi hybrid models and technologies Start & stop systems equip most of new Audi cars today, and the firm will extend its range of hybrid vehicles over the coming years. The programme e-tron was born in 2009 at the International Auto Show in Frankfurt with the unveiling of the Audi e-tron, an R8 look-alike EV sportscar. Ever since, other e-tron models of existing cars have been designed, such as the A1 e-tron, A3 e-tron and R8 e-tron which will soon start a small-batch production. A short individual presentation of the current Audi Plug-In Hybrid models and concept studies follow.

A1 e-tron with Range-Extender Table 1.5 shows some of the main characteristics of this electric vehicle equipped with a range extender.

Production status Under production Hybrid type EV with Range-Extender Hybrid architecture Series-hybrid Electrical machine 75 kW ICE Wankel engine:15 kW Battery Lithium-Ion:12 kWh EV Range 50 km

Figure 1.15: Audi A1 e-tron. Table 1.5: A1 e-tron technical information.

A3 e-tron The technical details of this newly introduced Plug-In Hybrid by Audi are gathered in Table 1.6.

Production status Under production Hybrid type PHEV Hybrid architecture Parallel-hybrid Electrical machine 75 kW/330 Nm ICE 1.4 TFSI:110 kW/250 Nm Battery Lithium-Ion:8.8 kWh EV Range 50 km

Figure 1.16: e-tron. Table 1.6: A3 e-tron technical information.

17 CHAPTER 1. INTRODUCTION 1.2. HYBRID TECHNOLOGY AT AUDI

A6 L e-tron Designed to promote electromobility for the Chinese market, the A6 L e-tron concept was revealed during the Beijing motorshow 2012. Its characteristics are given by Table 1.7. Note that Audi did not release any specification regarding the battery capacity.

Production status Concept Hybrid type PHEV Hybrid architecture Parallel-hybrid Electrical machine 70 kW ICE 2.0 TFSI:155 kW Battery Lithium-Ion EV Range 80 km

Figure 1.17: Audi A6 L. Table 1.7: A6 L e-tron concept technical infor- mation.

Sport Quattro Concept Presented at the Frankfurt Autoshow IAA 2013, the Audi Sport Quattro concept shows tremendous perfor- mance (see Table 1.8) due to the intelligent coupling between a powerful ICE and the electric machine.

Production status Concept Hybrid type PHEV Hybrid architecture Parallel-hybrid Electrical machine 100 kW/400 Nm ICE 4.0 V8 TFSI:412 kW/700 Nm Battery Lithium-Ion:14.1 kWh EV Range 50 km

Figure 1.18: Audi Sport Quattro Concept. Table 1.8: Sport Quattro Concept technical information.

Start & Stop and Drive situation identification The Start & Stop technology stops the engine when no contribution from the ICE is necessary, and starts it again as soon as power is required, as explained earlier in this report (see subsubsection 1.1.3.2). The Start & Stop complexity level has increased over the past years, with add-on technologies such as the Drive Situation Identification, developed at Audi by the EA-242 team. Originally named «Fahrsituationserkenung f¨urStart Stopp 2.0» [21], the feature aims at reducing the unpleasant short stops during which the engine goes off and on again less than 10 seconds later. The algorithm uses a deterministic approach which consists of identifying the current drive situation based on the past speed history, more particularly the stop times.

EV mode The Audi Hybrid cars are equipped with an EV mode, that can be activated by a simple push on the «EV» button located on the center console, if the battery state of charge allows it. In this mode, the ICE on/off power limits are higher so that the engine will start less frequently than in a regular hybrid mode. This increases the electric driving but depletes the battery faster.

Short Range Predictions Using very precise GPS data, short-range information about the oncoming topography and changes on the

18 CHAPTER 1. INTRODUCTION 1.2. HYBRID TECHNOLOGY AT AUDI

route can be computed. These details pertain to curves, gradients (up-/downhill) or speed limitations for instance and will be used by the control units to prepare the car accordingly. This technology enables an enhanced driving experience, as the car would react on time and with an appropriate behaviour to the route. The gear could be for instance properly set just before entering a curve, while improving the steerability and driving stability in the bend.

Basic drive strategy The regular drive strategy pertains to the management of the energy between the wheels, the electrical machine and the ICE. The ICE on/off power thresholds are determined by the drive strategy, as well as the amount of regenerative braking and the torque dedicated to the battery charge. The drive strategy can be improved by taking into account predictive data or analysing the driving situation.

Predictive drive strategy Long-range predictions enable forecasts in terms of power over a predefined route. These predictions can then be used to compute the fuel and electrical consumptions of a PHEV, to eventually determine an optimal succession of drive modes in order to reduce the fuel consumption and fully take advantage of the battery. The predictive drive strategy is the topic of this master thesis and will be explained in depth in the next chapter.

19 20 Chapter 2

Towards the optimal PDS: introduction

2.1 Principle

A drive strategy is a chosen set of energy management modes to be applied on the hybrid drivetrain, depending on some internal and external parameters, in order to reach the following specific goals as identified by Audi:

• achieve a reduction of fuel consumption

• enhance the EV-experience

• improve the comfort

• remove the need for a manual switch of mode while driving

The term predictive drive strategy (sometimes shortened to PDS) is used when the creation of this drive strategy involves data coming from route and energy needs predictions. This technology aims at reducing the fuel consumption by using an intelligent combination of drive modes on a given route [23]. The generic drive modes available for a Plug-In Hybrid are:

• EV: all electric through the whole power range of the electrical machine (40 kW power demand for electrical machine equipping the Audi A6 Hybrid).

• HEV: all electric up to a pre-defined limit (usually around 20 kW power demand).

• Charge: using load point shifting/increasing as much as possible so that the engine will be operated at its best operating points and possibly preventing the engine from stopping when the Start & Stop system would normally act, the charge mode aims at producing more power than needed in order to charge the battery.

To be able to determine how accurate each drive mode is, the route has to be known beforehand. The user simply enters the destination in the navigation system, which computes the route. The onboard computer will then analyse the route profile [29] and draw forecasts about some energetic characteristics. By doing this, anticipative information about the engine speed and torque demand [23] can be generated, which are called long range predictions. GPS data are provided by a cloud of 32 satellites enabling a computing of the car position at anytime, almost anywhere, with a precision of some meters. By combining these data with a well-documented real-time updated cartography, it is possible to get further details about the road, such as the curve radii, traffic information and speed limitations. Such information are already used in Adaptive Cruise Control (ACC) technologies which aims at anticipating the car’s short-term dynamic needs, like gear shift or suspension and steering adjustment. Some technologies are already dealing with the reduction of fuel consumptions on traditional ICE powered vehicles by accurately choosing the gear with help of GPS-based predictions (see oncoming subsubsection 2.2.2.1). As

21 CHAPTER 2. TOWARDS THE OPTIMAL PDS: INTRODUCTION 2.2. STATE OF THE ART

there are more degrees of freedom for a PHEV, considering both the electric and conventional drivetrains, the gains are theoretically much higher too. Based on these data, the route can be split into a certain amount of parts. Using recognition algorithms that will be explained later in this report, the onboard computer will determine an appropriate segmentation. For each segment, the expected electric and fuel consumptions will be calculated for all available modes (EV, HEV, Charge). The last step is to give this information to an optimizer that will create the optimal combination of modes to be applied on the route. Once the drive strategy has been computed, it will be implemented by the energy management system connected to both the electric propulsion system and the ECU. If some modifications concerning the computed itinerary occur while driving - such as an update of the traffic conditions or a diversion followed by the driver, the same loop can be repeated again and a new drive strategy will be calculated.

The master thesis The topic of the master thesis is the «Optimization of a predictive drive strategy for a Plug-In Hybrid vehicle»1. The targets are the following:

• Identify the different steps towards the implementation of a predictive strategy in a Plug-In Hybrid Audi. • Develop a tool capable of simulating the general operation of the predictive drive strategy algorithm. • Analyse the challenges and limits of pilot programs and algorithms.

Due to the high level of confidentiality, information related to the planning of the Predictive Drive Strategy project is very limited. This technology is expected to equip the next generation Plug-In Hybrid vehicles of Audi within three years.

A total of 7 engineers have been attached to the project:

• 1 person from the team I/EA-211 (Concept development Electric Drivetrain), in charge of the fuel and energy consumption simulations • 1 person from I/EA-222 (Powertrain functions and applications) dealing with long-range predictions • 2 persons from I/EA-252 (Software Development/Drivetrain functions) working on the analysis of the long-range predictions • 3 persons from the I/EA-242 supervising the project as well as focusing on measurements and strategy optimisation

2.2 State of the art 2.2.1 Research studies In the field of predictive drive strategies, the PhD dissertation of Michael Back [23] from the University of Karlsruhe is a reference. This analytic report was issued in 2005 and includes several references to previous works in the area. It covers the hybrid variants, different technologies to generate predictive data, method of determining the drive strategy based on the non-linear discrete car model and their implementation in a parallel hybrid model. Note that the non-linearity comes from the gear shift. Predictions are the base of an anticipated drive strategy. The route predictions come from:

• navigation system • up-/downhills and slopes • speed limitations • curves

1Original title in German: «Optimierung der vorausschauenden Antriebssteuerung bei einem Plug-In Hybrid».

22 CHAPTER 2. TOWARDS THE OPTIMAL PDS: INTRODUCTION 2.2. STATE OF THE ART

• driver’s behavior (reaction to speed limits and in curves for instance), and realtime informations (traffic info via RDS-TMC).

Three types of sensors can help in determining these predictive data:

• distance sensors: assess the traffic density by measuring an average car-to-car distance, gather information as to how the driver accepts to following a slower vehicle • cameras: read road signs, assess traffic density • weather sensors: evaluate driver behaviour with rain, take into account weather depending speed limita- tions

For further optimizations, a detection of traffic lights, stop signs, crosswalks and intersections can be con- sidered. From all this route information, the acceleration and speed profiles are determined. The required torque and engine speed are then computed by sending the speed and acceleration profiles through the physical car model, while the driver’s torque and power demands are known. This information is then given to the optimizer, which focuses on fuel-consumption reduction. The optimizer works with discrete data and is based on dynamic programming instead of the search of a global optimum. Dynamic programming reduces the original problem to a smaller, more limited sub-problem. In other words, a gliding window will define the predictive data that the algorithm has to consider in order to solve the problem. In Back’s report, this window can be up to 65 seconds wide. At a time t0, only the predictions between t0 and t0 + 65s will be considered to optimize the fuel consumption in this window and use the electrical energy accurately. Based on the oncoming route profile contained in the sliding window, predictions about the torque and power needs will be considered. For each optimization, a defined share of the battery capacity will be consumed. The electrical machine is used to support the ICE (in uphills for instance) while respecting this allowed electrical energy consumption. For some examples including an uphill followed by a downhill, this energy share can be equal to zero as the regenerative braking will compensate the energy used on the uphill. However, in most of the common cases, the amount is likely to be strictly positive to allow a visible improvement in the fuel consumption. The result of this optimization is a set of regulation commands for the ICE, the electrical machine, the gearbox and the brakes. The whole algorithm is then integrated in a car model and assessed in an experimental environment, with real speed profiles. For all tested configurations, the resulting fuel consumption is lower with a predictive drive strategy than with a conventional hybrid strategy. The author concludes that due to the high variability among the tests, more simulations are needed to provide statistic average fuel savings. Simulation conflicts in the full-car model arise from the gear box, as gear shifts are often suggested by the predictive drive functions and not implemented by the gear control unit. With a simple predictive drive strategy, the simulations have shown that a reduction of at least 2% is possible. Many technical resources at Audi refer to Back’s «Pr¨adiktive Antriebsregelung zum energieoptimalen Betrieb von Hybridfahrzeugen». However, the approach implemented by the firm differs from it on several aspects, as described in the chapter3.

2.2.2 In the industry 2.2.2.1 BMW: Predictive Power Management In 2012, BMW had announced its new Predictive Power Management feature [30]. A range of GPS-based functions supports this new functionality by allowing an anticipative control of the gearshift. The road forecasts include a consideration of the up-/downhills, the curves as well as other road obstacles such as roundabout. The Predictive Power Management is coupled to a panel of features to enhance the driver’s experience:

• Pure Dynamics: In Sport Mode, the car will adopt a dynamics and performance oriented behaviour. Based on the road forecasts, an optimized gearshift is implemented to make the right torque available at the right moment, such as at the end of a curve.

23 CHAPTER 2. TOWARDS THE OPTIMAL PDS: INTRODUCTION 2.2. STATE OF THE ART

• ECO PRO mode features – ECO PRO Foresight Assistant: this functionality gives the driver tips to reduce the fuel consumption based on oncoming speed reductions, obstacles or possible use of coasting. – ECO PRO Coasting: the engine can be decoupled when the driver releases the gas pedal (following an indication from the Foresight Assistant for instance), in a range of speed of 50 km/h up to 160 km/h. BMW announces fuel savings that can reach 5% with this feature. – ECO PRO Route: the navigation systems allows the driver to choose between direct route and more economical ones to enhance fuel savings, while considering the driver’s driving behaviour. – ECO PRO Analyser: simply analyses the driving behaviour and rates its level of eco-friendliness. – ECO PRO Efficiencytainment: overall interface that displays which of the features are currently being used. The engine temperature regulation is also based on the Predictive Power Management. These features are gathered under the ActiveHybrid technology package.

2.2.2.2 BMW: Anticipatory Energy Management Already mentioned in the 2012 media release [30] as well, the use of Predictive Power Management to optimize the hybrid driving strategy has been confirmed at the 22nd Aachen Colloquium Automobile and Engine Technology 20132. Public information about this technology is rather rare but BMW has already acknowledged that the Antic- ipatory Energy Management system will equip the BMW i8 [31]. Just like the Predictive Power Management, the Anticipatory Energy Management considers the road profile, as well as up-/downhills, curve radii and speed limits to compute energy consumption forecasts for both the ICE and the electric drivetrain. In a non-anticipated strategy mode, the vehicle would run electrically in cross-country and city areas. How- ever, the Anticipatory Energy Management feature will allow an intelligent management of the drive modes meaning that pure electric driving will occur where the efficiency is optimal. The document [31] does not stipulate whether the Anticipatory Energy Management calculates an optimal driving strategy. It seems that a priority based optimization is used (see section 3.6 about the priorities) to ensure an EV driving for the last section of the itinerary3.

2.2.2.3 Daimler: predictive drive strategy for E-trucks In 2011 two members the Institut f¨urSystemdynamik of the University of Stuttgart had published a short article describing a project of drive strategy for hybrid trucks carried out in cooperation with Daimler [33]. Using predictions regarding the route topography and the speed limits, commands will be sent to the hybrid powertrain in order to reduce the fuel consumption, with a considerable energy input from the regenerative braking. The optimization is based on two different algorithms: • a short-range non-linear optimizer which defines the gearshift and engine speed. • a long-range linear optimizer which analyses the route, classifies its parts into different sections depending on their gradient profile and then determines progression of SOC of the battery. As no further publication than [33] has been issued on this project by the authors so far, one can assume that the study is either still ongoing or protected by Daimler.

27 - 10 October 2013 in Aachen, Germany. 3Similar to an EV - coming home function developed by Audi [32].

24 Chapter 3

Towards the optimal PDS: methods

After a description of the measurement and simulation processes, the methods used for each step of the calcu- lation of a predictive drive strategy will be developed.

3.1 Measurements and datasets 3.1.1 Test car: Audi A6 Hybrid Several measurement series have been performed using the department’s A6 Hybrid with a large-capacity battery in different driving environments, such as city, city traffic, motorway and country roads, in the region or on Audi’s test track. This parallel hybrid is equipped with an onboard measurement unit recording one point (set of parameters) each ∆t = 10−2 s for most parameters1. A laptop connected to the car’s control unit allows to set specifically the driving mode. The main drive strategy parameters can be directly set through the software user interface:

• power above which the ICE will start • power under which the ICE will stop

By changing these parameters, the following driving modes can be implemented live:

• EV mode: the ICE starts and stops as seldom as possible, i.e only when the required power exceeds the electrical machine performances. The battery will therefore deplete faster than in the other modes. • HEV mode: the ICE start and stop limits are lowered and is mostly used to hold the battery charge. The electrical machine mainly powers the car at low speed and is then supported by the ICE for greater accelerations. A load point optimization is also performed by the ECU to briefly charge the battery with the ICE when possible. • Charge mode: the ICE runs most of the time, at a higher engine speed some of the time. The aim is to produce more torque than needed at the wheels in order to charge the battery.

Table 3.1 shows the main characteristics of the modified Audi A6 Hybrid used for the different calculations.

3.1.2 Measured parameters While the pilot drives on the selected itineraries between Ingolstadt and Munich, a copilot ensures that the appropriate driving mode has been correctly set in the ECU and monitors the measurements. All of these measurements have then been gathered and categorized into 28 sections:

• 3 motorway sections (2 driven in HEV mode, 1 in EV mode)

1The cumulated injected volume of fuel is saved every 10−1 s.

25 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.1. MEASUREMENTS AND DATASETS

Table 3.1: Audi A6 Hybrid Technical Data[34].

Mass empty m 2000 kg Aerodynamic drag coefficient Cx 0.26 Frontal Area A 2.3 m2 Rolling resistance coefficient fr 0.0095 Wheel radius rwheel 339.65 mm ICE max. power 155 kW Final drive reduction ratio Ufinal 3.076 st 1 gear ratio U1 4.714 nd 2 gear ratio U2 3.143 rd 3 gear ratio U3 2.106 th 4 gear ratio U4 1.667 th 5 gear ratio U5 1.285 th 6 gear ratio U6 1.000 th 7 gear ratio U7 0.839 th 8 gear ratio U8 0.667 Electrical machine power 40 kW Battery capacity 10.678 kWh

• 14 country road sections (9 driven in HEV mode, 5 in EV mode) • 11 city sections (6 driven in HEV mode, 5 in EV mode)

In each set of data, a thorough panel of values has been recorded by the onboard measurement unit. The most useful parameters are given in Table 3.2.

Table 3.2: Measured parameters, units and signal names.

Parameter Symbol Unit Signal name Sync. time Vehicle speed v km/h Eng v VW 10 ms Engine torque Tengine N.m Pt tqClu VW 10 ms Engine speed nengine rpm Pt nClu VW 10 ms Gear number gn - Tra noGear VW 100 ms ICE on/off - - CoVOM bHybDrv VW 10 ms SOC SOC % HvBatt rSoc VW 10 ms Cumulated fuel consumption - L vtrip w 100 ms

In parallel to the measurements related to the master thesis, a large test campaign on the A6 Hybrid has been conducted (1 day a week in average) on Audi test area (Pr¨ufgel¨andeNeustadt). Based on two typical city rounds, one coming from Ingolstadt and the other one from Beijing, a panel of 13 different electrical strategies have been tested. Even though these tests were not directly related to a drive strategy prediction, they pertain to a practical research of the optimal drive strategy in a urban environment. Each one of the two reference rounds was characterized by a given speed profile. This profile was then reproduced precisely on a circular test track with the instrumented A6 Hybrid. The strategies differs from each other on the following points:

• speed above which the ICE starts • speed below which the ICE stops

26 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.2. SIMULINK MODEL

3.2 Simulink model

A full Simulink model has been developed for an Audi PHEV e-tron project in order to virtually test the vehicle on some speed profiles such as standard driving cycles and assess the levels of fuel consumptions and emissions. This model has been adapted to the characteristics of the modified Audi A6 Hybrid. The model is composed of 8 main blocks:

• Speed profile selection block: the user can choose the cycle to be simulated in this unit (see subsection 3.2.1). • Resistance forces block: the rolling and aerodynamic resistances are computed in this block, based on well-known equations of movement, as detailed in subsection 3.2.2. • Gear box block: simulating the 8-gear tiptronic automatic box described in subsection 3.2.3. • Drive strategy block: representing the Energy Management System described in 3.2.4, it decides of the Start & Stop, regenerative braking and charge mode parameters. • HV Battery/Boardnet junction block: computing the energetic transfers between boardnet and high- voltage battery through an interpolation based on the engine speed and torque. • HV Battery block: simulating the HV battery charge and discharge. • 2 blocks for the consumption and emission calculations: compute the fuel consumption as well as the NOx emission levels (see subsection 3.2.5).

Some of the main blocks will be further detailed below.

3.2.1 Input: road profiles Several driving cycles have been designed over the past years to constitute a common base for the consumption and CO2 emission evaluations. The most widespread and used in the world are the NEDC (New European Driving Cycle), the FTP-75 and the World-Harmonized Light-Duty Vehicles Test Procedure (WLTP). These cycles are parametrized and imported in the Simulink model to serve as input for the speed profile.

NEDC The New European Driving Cycle has been introduced in the 90’s as the combination of two preexisting cycles: four ECE-15 Urban Driving Cycles followed by one Extra Urban Driving Cycle. It is nowadays used in Europe as the reference speed profile to check the emission compliance of passenger cars [35]. The speed profile is given in Figure 3.1. The overall length slightly exceeds 11 km for a duration of 1180 seconds.

Figure 3.1: Speed profile of the NEDC.

27 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.2. SIMULINK MODEL

FTP-75 The FTP-75 is a US city driving cycle implemented by the Environment Protection Agency since 1978, aiming at providing a common base for passenger car emission level assessment [35]. It is composed of two phases, a cold start phase and a transient one, simulating a total of 23 stops as shown of Figure 3.2. The FTP-75 cycle is 17.77 km long and lasts 1874 seconds.

Figure 3.2: Speed profile of the FTP-75.

World-Harmonized Light-Duty Vehicles Test Procedure Abbreviated in WLTP, this new set of test procedures has been developed by experts from the European Union, Japan and India, and adopted by United Nations Working Party on Pollution and Energy (GRPE) [36]. As mentioned on the agenda of the next World Forum for Harmonization of Vehicle Regulations [37], which was held in Geneva in March 2014, the proposal for draft of the WLTP was still to be approved. If so, this set of procedures is likely to replace the current NEDC very soon to assess the compliance towards Euro 6. Based on the vehicle class, depending on the power-to-mass ratio and top speed, one of the three developed driving cycles will be chosen. For most of the regular passenger cars, the class 3 driving cycle will be used (power-to-mass ratio greater than 34 kW/ton and top speed over 120 km/h). Its length is 1800 seconds and is shaped as shown by Figure 3.3.

Figure 3.3: Speed profile of the WLTP [36].

28 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.2. SIMULINK MODEL

3.2.2 Power calculation Because power calculation is the base of the data processing to eventually determine the drive strategy, a physical explanation of how to derive the power from the measurements will be given.

Figure 3.4: Representation of the forces applied on a car [38].

The direction of motion is x and ~ex is the corresponding unit vector. Its perpendicular is oriented along the z-axis and has ~ez as unit vector. As shown on Figure 3.4, when considering the whole car as a system, the external forces applied to it during acceleration on a slope of angle α are:

• the weight: W~ = m × ~g = m × g × (− cos (α) × ~ez − sin (α) × ~ex) • the total ground reaction: N~ = N~1 + N~2 = N × ~ez • the total rolling resistance for both axles: F~r = F~rf + F~rb = −m × g × fr × ~ex ~ 1 2 • the aerodynamic resistance: FL = − 2 × Cx × A × ρ × x˙ × ~ex • the net tractive force: F~x = Fx × ~ex

Taking into account the inertia contribution of the rotating parts in an equivalent mass mj, the equation of motion along the direction of motion ~ex can be written as follow:

(m + mj) × ~a = F~x + W~ + F~r + F~L + N~ 1 (3.1) (m + m ) × x¨ = F − m × g × sin (α) − m × g × f − × Cx × A × ρ × x˙ 2 j x r 2

The power required at wheels is Px the power of the net tractive force Fx:

Px = F~x · x˙

= (Fx × ~ex) · (x ˙ × ~ex) (3.2)

= Fx × x˙

29 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.2. SIMULINK MODEL

Fx can be expressed using (3.1) above. By introducing this expression into (3.2), the following equation is obtained: 1 P = (m + m ) × x¨ × x˙ + m × g × (sin (α) + f ) × x˙ + × Cx × A × ρ × x˙ 3 (3.3) x j r 2 A discrete version of this equation is given in equation (3.4), which is convenient when working with the   measured speed X˙ = X˙ 1 X˙ 2 X˙ 3 ... X˙ n . In the measurement data, the vector Eng v VW represents the vehicle speed and is measured every ∆t = 10 ms.

Px,i = Fx,i × X˙ i ∀i ∈ [1; n − 1] X˙ − X˙ 1 (3.4) = (m + m ) × i+1 i × X˙ + m × g × (sin (α) + f ) × ×X˙ + × Cx × A × ρ × X˙ 3 j,i ∆t i r i 2 i

If the equivalent mass mj was not originally considered in the Simulink model, it can be taken into account when computing the power from measurement data with equation (3.4). To do this, the gear speed must be know for every measurement point. In the dataset, this task is carried out by the 100 ms time synchronous vector Tra noGear VW, as previously shown on Table 3.2. Each inertia of the rotating parts Ji can be expressed through the equivalent mass mj using the formula (3.5)[38].

ΣJ m = i j r2 wheel (3.5) Jw + Jp + Je = 2 rwheel The main rotating parts influencing the equivalent mass are:

2 • the wheels and axles, polar inertia Iw ≈ 2.7 kg.m : Jw = Iw 2 2 • the propeller shaft, polar inertia Ip ≈ 0.04 kg.m : Jp = Ip × Ufinal 2 2 • the engine, flywheel and clutch, polar inertia Ie ≈ 0.4 kg.m : Je = Ie × (UfinalUi)

In the end, the equivalent mass mj can be estimated for each gear as sum of the aforementioned inertias divided by the square of the wheel radius, as shown on the mathematical expression (3.5). Approximations for each gear are given in Table 3.3, where it can be noticed that the influence of the equivalent mass is much bigger for small gears. Figure 3.5 also shows the evolution of the ratio between equivalent mass and total mass as function of the total gear ratio.

Table 3.3: Approximations of the equivalent masses for each gear of the 8-speed gearbox.

Gear 1st 2nd 3rd 4th 5th 6th 7th 8th mj [kg] 756 351 172 118 81 59. 50 41 mj/m [%] 37.8 17.6 8.6 5.9 4.1 3.0 2.5 2.1

3.2.3 Gearbox The Audi A6 Hybrid is equipped with an 8-speed tiptronic gearbox, offering a semi-manual mode at the steering- wheel. The gear ratios for the 8-speed tiptronic gearbox are given in Table 3.1. Based on the torque demanded by the driver, the gearbox module in Simulink will select the appropriate gear among the 8 available. As this module directly works with vehicle speed in km/h and engine speed in rpm, 0 the gear ratios Ui for i ∈ [|1; 8|] entered in the program are actually the original gear ratios Ui, multiplied by a conversion coefficient, as shown in (3.6).

30 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.2. SIMULINK MODEL

Figure 3.5: Ratio between the equivalent mass of the rotating parts and the empty mass of the A6, as function of the total gear ratio.

0 nengine [rpm] Ui = vcar [km/h] n [rpm] n [rad/s] ω [rad/s] v [m/s] = engine × engine × wheel × car (3.6) nengine [rad/s] ωwheel [rad/s] vcar [m/s] vcar [km/h] 60 1 1 = × UiUfinal × × 2π rwheel 3.6

0 The resulting extended gear ratios Ui are given in Table 3.4. Table 3.4: Simulink model extended gear ratios.

0 0 0 0 0 0 0 0 U1 U2 U3 U4 U5 U6 U7 U8 113.2 75.50 50.59 40.05 30.87 24.02 20.16 16.02

3.2.4 Energy Management System The EMS is the unit in charge of managing the energy strategy in the car. The Start & Stop function is implemented by this unit, as well as the drive strategies. Power/engine speed tables for the Start & Stop feature are integrated in the software and determine the engine on and off state.

3.2.5 Consumption and emissions The consumption calculation is based on injected fuel mass maps built on the torque and engine speed. As mentioned previously, one measurement point is created each ∆t = 10−2 s. For each point, the mass of fuel injected in the cylinder is known and can then be integrated over the whole dataset, and converted to energy or volume. A similar procedure is used for the NOx emission levels.

31 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.3. ENERGY VS POWER DISTRIBUTION

The overall distance of pure electric drive is also computed by this unit using an integration of the binary motor on/off vector (1 if the ICE is on, 0 if ICE off).

3.3 Energy vs Power distribution

For practical reasons, all the information needed for the predictive drive strategy should be gathered in one object type. This type should allow a split of the route in recognized segments and the computing of consumption estimates, while being easily understandable from a physical point of view. The first idea to be considered is the removal of the time dimension in the predictive data object type. The reason is that time implies a broad range of information which is here not needed and can sometimes make it more difficult for the application. Assuming that the goal is to compare two power profiles to determine if they belong to the same road type or not, there might already be some issues if time is involved. Indeed, it is not obvious that when predicting the power profile of a given road from A to B in one direction and comparing these predictions with a second simulation in the other direction (from B to A), the algorithm would recognize that both belong to the same road. The second reason is to model all types of predictions according to one single pattern. For algorithmic use, it is absolutely necessary to format all data hence forbidding random vector sizes for instance. The energy distribution as function of accurately determined power intervals meets both aforementioned requirements, and is fast and easy to compute from energy and power vs time profiles. Once the power has been computed from the prediction or measured data using equation (3.4), a local energy estimate can be computed for the ∆t-spaced n-point measurement interval:

∆iE = |Px,i| × ∆t ∀i ∈ [1; n − 1] (3.7)

All the points whose power values P are included in a given power interval P− ≤ P ≤ P+ are grouped in a set I. The energy increments of all the elements of I will then be computed according to equation (3.7) and summed altogether to form the energy bar EP−≤P ≤P+ . Mathematically, this means:

X EP−≤P ≤P+ = ∆kE k∈I X = Px,k × ∆t (3.8) k∈I X = ∆t × Px,k k∈I

By structuring joint and accurate power intervals, a meaningful graphical representation of the energetic profile of a route can be produced. The number of power intervals on the x-axis needs to be both precise enough to allow flexibility - especially for consumption calculations (see section 3.5) - and limited to avoid an unnecessary complexity. The limits between the intervals represent physical boundaries, either between separate drivetrain systems - such as the ICE and the electrical machine -, or within some units - such as the regenerative braking system. This allows, as mentioned previously, a flexibility for the data treatment. The energy distribution histogram will be split into 14 intervals, covering the whole power range of the drivetrain. The generic configuration is given by Table 3.5. The energy is presented per 100 km to facilitate the reading of such a diagram. To summarize the energy vs power histogram process, the user first gives an itinerary through the navigation system, then long-range predictions estimate the speed profile. The power and energy increment profiles are computed using equations (3.4) and (3.7), and the energy vs power distribution is finally computed following the power intervals described previously. Figure 3.6 shows the energy vs power histogram for the New European Driving Cycle and the FTP-75 for the A6 hybrid.

32 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.4. SEGMENTATION

Table 3.5: Intervals for the energy vs power histograms.

Interval # 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Lower bound −∞ −Preg 0 kW PEM,max PICE,max A6 Hybrid values −∞ -40 kW 0 kW 40 kW 155 kW Upper bound −Preg 0 kW PEM,max PICE,max PICE+EM,max A6 Hybrid values -40 kW 0 kW 40 kW 155 kW +∞

−Preg maximum power at wheels that can be fully recuperated using regenerative braking 10 bars are equally spaced between 0 kW and PEM,max PEM,max electrical machine maximum output power at wheels PICE,max ICE maximum output power at wheels PICE+EM,max ICE and electrical machine (E-boost) maximum output power at wheels

(a) (b)

Figure 3.6: Energy distribution [kWh/100km] vs power [kW] for the NEDC cycle (a) and the FTP-75 cycle, for an Audi A6.

3.4 Segmentation

During the real onboard process, the second step after the calculation of predictions will be the virtual split of the route into segments, according to their internal properties. The aim of this stage is to somehow gather the similar environment and give the optimizer a limited amount of mode combinations. As the computing time can rapidly grow with the number of degrees of freedom, it is not reasonable to suggest a drive mode change at any time. A drive mode will be defined for the whole segment and will not be changed before reaching the end of it, except in special cases. A route segmentation is already used in the ACC technology (Adaptive Cruise Control) which aims at automatically maintaining a certain distance with the traffic ahead. To enhance the comfort when the ACC is activated, the article [39] suggests the segmentation of the route in categories in order to adjust the distance. The categories are fast road, country road, urban road and exit. Depending on the road type, the distance to be kept with the vehicle ahead will vary. In city traffic, the lower average speed allows a shorter security distance, hence a need to adapt the ACC parameters. In his work about anticipative assistance systems, Michael Schraut [29] mentions 8 categories of roads, characterized by a range of speeds. For the predictive drive strategy, segments will be compared to determine their type. The similarities between two segments can be assessed by several internal parameters and functions. Two methods are presented below: one based on the Kolmogorov-Smirnov test function, and one based on a statistical study of the energy distribution.

33 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.4. SEGMENTATION

Moreover, the question of the segment classification is a central topic. The segment comparison can be performed within the given route, without any reference categories. In this binary approach, one segment of the route will be compared with another segment of the route. As a result, only the fact that two segments are similar or different will be assessed. This is the background of the Kolmogorov-Smirnov test function. However, as soon as a precise category identification is required, reference segment categories have to be involved in the process and appear as input. This approach is used with the second method tested here, based on the statistical parameter study.

Figure 3.7: Global segmentation process.

The global segmentation process is given on Figure 3.7. For both tested methods, the route is first split into sections of equal lengths (500 m for instance). The sections are then compared with each other using the power cumulative distribution (Kolmogorov-Smirnov) or compared with references (statistical parameter study) considering their energy vs power distribution.

3.4.1 Kolmogorov-Smirnov test In the first solution implemented, Matlab’s function kstest2.m was used. This function is based on the two- sample Kolmogorov-Smirnov test, aiming at assessing whether the two compared signals come from the same continuous distribution.

Theory Two temporal signals s1 and s2, not necessarily sharing the same time-base, represent the power at the wheels for two given segments. The idea is to determine whether the energy distribution as function of the power (see an example with the NEDC and FTP-75 on Figure 3.6) have similarities. As this method relies on the power distribution functions and not directly the energy, the aim is to determine if the frequency at which the power values appear in the signal can characterise it. The Kolmogorov-Smirnov test function will first compute the distributions S1 and S2 of the signals s1 and s2, and then calculate the Cumulative Distribution Functions (CDF) F1 and F2:

F1 (x) = P (s1 ≤ x) (3.9) F2 (x) = P (s2 ≤ x)

34 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.4. SEGMENTATION

P represents here the function «probability» which calculates the probability (or the number of occurrences in this case) that the signal takes values smaller than x. The method eventually gives a decision concerning the non-respect of the null hypothesis. The null hypothesis means here that the two signals stem from the same continuous density. Consequently, the function kstest2.m will return 0 if the null hypothesis is not rejected, or 1 if the null hypothesis is rejected. In a nutshell, 0 means that p1 and p2 share the same continuous distribution. To make this decision, the function will compare the CDF F1 and F2 and remember the biggest absolute difference, called Kolmogorov-Smirnov statistic D1,2:

D1,2 = sup |F1 (x) − F2 (x)| (3.10) x

The Kolmogorov-Smirnov statistic will finally be compared with a parameter Dlim (α) calculated for large samples using the alpha-value α, which defines the comparison’s similarity threshold, and the sizes n1 and n2 of the signals s1 and s2: r n1 + n2 Dlim (α) = c (α) × (3.11) n1 × n2 c (α) is a coefficient given by Table 3.6.

Table 3.6: Values for c (α).

α 0.10 0.05 0.025 0.01 0.005 0.001 c (α) 1.22 1.36 1.48 1.63 1.73 1.95

If D1,2 > Dlim (α) then the null hypothesis is rejected for the given alpha level, meaning that the function will return 1 and that s1 and s2 do not belong to the same distribution. On the other hand, a value of 0 might imply some similarities in between the segments.

3.4.2 Statistical parametric comparison In a second type of comparison, the choice was made to test if some road types/environments could be recognized. This necessarily implies that reference types will be added as input for the function. Among the 28 sections from the measurements databank, some pertaining to city driving and country roads are selected and cut into 500 m segments, resulting in a total of 152 segments (115 country road segments and 37 city segments). For each segment, the energy distribution is computed, as explained in section 3.3, and stored. At the end of this stage, two datasets of energy distribution are available: one with 37 energy distributions of city driving segments, and one with 115 energy distributions of country roads. The integer interval K = [|1; 14|] is the indexing of the 14 power intervals used to compute the energy vs power distribution. A large range of 63 statistical parameters are calculated for all energy distributions of those two sets.:

1 2 1 • variance σ (E) = card Σ Ek − E , where E = card Σ Ek K k∈K K k∈K

• maximum energy Emax = Ekmax = maxEk k∈K

• power of the maximum energy P (Emax) = {Pm |Em = Emax }

• minimum power Pmin = inf P, where P = {Pk |Ek 6= 0 , k ∈ K }

• maximum power Pmax = sup P

• power range of the energy distribution ∆P = Pmax − Pmin

35 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.5. FUEL AND ELECTRICAL CONSUMPTIONS

  k • power PΣEi

3.5 Fuel and electrical consumptions

Once the route has been calculated and segmented into a certain number of sections as described previously in section 3.4, the aim is now to estimate the fuel and electrical consumptions of each section, for the three

36 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.5. FUEL AND ELECTRICAL CONSUMPTIONS

Figure 3.8: Example of the three different types of intervals that can be found for a given parameter (called X) and two categories (Country road and City). available modes (EV, HEV and Charge, see 3.1.1). This calculation is based on the long-range prediction and will therefore use the energy vs power distributions. Each bar of an energy vs power distribution corresponds to a specific range of power. By summing the positive power bars (energy consumed) together and subtracting the negative power bars (energy recuperated), the total energy balance of the section is obtained. Note that the energies represented in an histogram are calculated at the wheels. The aim of this study about fuel and electrical consumptions is to evaluate how much fuel, respectively how much electrical energy will be taken out of the tank, respectively out of the battery. A certain conversion from the energies at the wheels is therefore needed to transform them into fuel and electrical consumptions. In this conversion, the following topics have to be considered:

• the efficiency chain of the drivetrain components: ICE, electrical machine, gearbox, battery and wires. • the regenerative braking settings: contribution of mechanical brake vs regenerative braking • the E-boost settings: contribution from the ICE vs electrical machine

As an example, the bar 4-8 kW in an energy vs power distribution of a given segment is considered for the EV-mode. This bar belongs to the range of the electrical machine operated in EV-mode. As a result, no contribution from the ICE is expected for this range of power between 4 kW and 8 kW and the consumed energy will be completely provided by the battery. The energy at wheels corresponding to this 4-8 kW power at wheels bar now needs to be converted to energy taken from the battery. To do this, the overall efficiency chain needs to be taken into account. When the electrical machine powers the wheels, the energy comes first from the battery, then goes through the electrical machine and finally through the gear box2 to end up in the wheels. The energy taken from the battery for this range of power is therefore the energy at wheels divided by the combined efficiency of the gearbox, electrical machine and the battery. This combined efficiency is necessary

2Due to the parallel hybrid configuration

37 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.5. FUEL AND ELECTRICAL CONSUMPTIONS

Figure 3.9: Process for statistical comparison.

below 1, which consequently implies that the energy taken out of the battery will be greater than the energy at wheels. The previous example shows two main requirements in the consumption computations:

• each range (bar of the histogram) of power has technically a different conversion ratio, due to a different operating point for the efficiency chain and the influence of regenerative braking or E-boost for some power values. As a result, each bar of the energy vs power distribution will have its own conversion coefficient. • each drive mode has different boundaries for the start and stop of the ICE. Consequently, the electrical fuel consumptions will not be based on the same ranges of powers (bars) for all drive modes.

The requirements mentioned above have been translated to mathematical and physical terms by introducing two parameters:

• a vector of 14 conversion parameters called (αi)i=1..14 for the electrical consumption and a similar vector for the fuel consumption called (βi)i=1..14. • the range of power to be considered for the calculation, which is different for fuel consumption and electrical consumption. This is expressed through the index i0 and i1 in the sum presented in equations (3.12) and (3.13).

The electrical consumption e in [kWh/100 km] of a given segment can be expressed as the sum of all converted energy bars Econv,i located between the index i0 and i1. A converted energy bar Econv,i is the product of the

38 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.6. PRIORITIES

original energy bar Ei and its conversion parameter αi.

i X1 e = Econv,i i=i 0 (3.12) i X1 = αi × Ei

i=i0 Similarly, the fuel consumption f in [kWh/100 km] of a given segment can be expressed as the sum of all converted energy bars Econv,i located between the index i0 and i1. A converted energy bar Econv,i is the product of the original energy bar Ei and its conversion parameter βi.

i X1 f = Econv,i i=i 0 (3.13) i X1 = βi × Ei

i=i0

Note that the index i0 and i1 have different values depending on the drive mode and the type of consumption to be calculated (electrical or fuel). Figure 3.10 shows an example of the bars to be considered in the electrical consumption (a,c) and fuel consumption (b,d), for the EV-mode (a,b) and HEV-mode (c,d).

3.6 Priorities

Among the three available driving modes, some will be more relevant than others depending on the driving environment. For instance, it might be more efficient to drive on highway without any electrical contribution since the ICE would run at good load points. As a result, a HEV or CHARGE mode would be preferred in that situation. The global idea here is about matching driving modes with driving environments through a priority system. This priority system could for instance be based on points. For each driving environment or segment, each mode would be granted a certain amount of points depending on how accurate it would be. The number of points given to each mode/environment association might be based on the fuel economies, the EV-experience, the comfort or some requirement regarding Zero Emission Zone (ZEZ). The priorities can be applied on the two types of segments presented in section 3.4, offering two types of prioritisation: • identified categories, such as city driving or motorway, referred to as «offline prioritization». • uncategorised segments coming from a segmentation process that does not rely on global categories (see statistical parametric comparison in 3.4), referred to as «online prioritization».

3.6.1 Offline prioritization This option assumes that the route has been segmented and its segments categorized in predefined driving environments. A table listing the priority modes is integrated to the drive strategy management unit before- hand. Then by simply comparing the segment category with the associated preferred modes, the optimizer can determine an enhanced strategy. Table 3.7 gives an example of how an offline priority table integrated in the car’s onboard energy management unit can look like. One major advantage of the offline prioritization is that it can take into account ZEZ and Low Emission Zones (LEZ). Some cities in Europe, like Stockholm or Berlin have Low Emission Zones (LEZ) in which heavy emitting vehicles are forbidden. Some Zero Emission Zone (ZEZ) projects are already under consideration. Therefore, if a ZEZ appears on the car’s itinerary, the main constraint will not be the reduction of the fuel consumption on the long range anymore. The driving factor will be to drive in full-electric mode through this

39 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.6. PRIORITIES

e: electrical consumption l: fuel consumption EV

(a) (b) HEV

(c) (d)

Figure 3.10: Example of values to be taken into account for the consumption computation (red bars): in EV mode, electrical (a) and fuel (b) - in HEV mode, electrical (c) and fuel (d).

Table 3.7: Example of offline priority table.

Category 1 Category 2 ... High priority mode EV HEV ... Medium priority mode HEV CHARGE ... Low priority mode CHARGE EV ...

specific area. This implies that the batteries must have a SOC level at the ZEZ entry allowing the vehicle to drive through and exit the area without starting the internal combustion engine. In case such an environmental zone is identified by the navigation system, it will be integrated to the segmentation as a special category. Consequently, the offline priority table will display an absolute priority for the EV mode when driving through a ZEZ. The issues raised by this approach relate to the possibility of associating offline one mode to one driving category. A piece of answer comes out of the several optimizations that have been run using the optimizer described later (see section 3.7). For the case where categories are represented by road types, it appears that the absolute optimal solution hardly ever associate the same mode to the same road type in 100% of the tests. This means that giving fixed offline priorities might be something risky. However, the distance between the

40 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.7. OPTIMIZER

offline-priority based strategy and the absolute optimal strategy needs to be studied to assess how the latter differs from it. Moreover, a strategy based on offline priorities might enhance the EV experience even if it does not necessary optimize the fuel consumption.

3.6.2 Online prioritization Unlike the offline method, the online prioritization computes in real-time the modes to be prioritized for each segment. In this approach, the segments have not been categorized, which means that the priorities must be established according to the segment’s characteristics given by its energy vs power distribution. As explained previously in section 3.5, the fuel and electrical consumptions can be computed using the energy vs power distribution. A comparison between the different consumptions can help setting priorities for the driving modes. The following example shows what parameters can be taken into account and how they can be handled to find in the end the driving mode to prioritize. The main parameters to consider are:

HEV mode - fuel consumption of segment i [kWh] fHEV,i HEV mode - electrical consumption of segment i [kWh] eHEV,i EV mode - fuel consumption of segment i [kWh] fEV,i EV mode - electrical consumption of segment i [kWh] eEV,i CHARGE mode - fuel consumption of segment i [kWh] fCHARGE,i CHARGE mode - electrical consumption of segment i [kWh] eCHARGE,X Assuming that the goal is to determine which of driving modes EV and CHARGE will be the most relevant to use on segment i, the following differences will help in making a decision:

Gain in fuel consumption between modes EV and CHARGE [kWh] ∆fCHARGE-EV,i = fCHARGE,i − fEV,i Gain in battery charge between modes EV and CHARGE [kWh] ∆eCHARGE-EV,i = eCHARGE,i − eEV,i

∆fCHARGE-EV,i By comparing the ratio RCHARGE-EV = with an established threshold value XCHARGE-EV, the ∆eCHARGE-EV,i driving mode to prioritize can be determined. The threshold value XCHARGE-EV sets the limit for RCHARGE-EV above which the EV mode is energetically more interesting to choose than the CHARGE mode. Physically, a high RCHARGE-EV ( XCHARGE-EV) means that the gain in fuel consumption in EV mode compared to the CHARGE mode is bigger that the gain in electrical consumption in CHARGE mode compared to EV mode. Consequently, the EV mode can be prioritized in such a situation. On the opposite, when RCHARGE-EV is low, the gain in battery charge in CHARGE mode is relatively predominant over the extra fuel consumed compared to the EV mode. This is a simple example of a method. Other parameters can of course be considered to determine an online priority list. Whatever the prioritization type, a priority-based optimization is expected to be much faster than a thorough fuel consumption optimization such as the one detailed in section 3.7. Finally, priorities could also be used as complementary information, and not as main optimizing parameter. This would amongst others favour some modes under given conditions. A classical case in the Zero Emission Zone example given previously. Relying only on priorities to compute the predictive drive strategy brings the risk that the strategy will not be the most fuel efficient. However, using priorities as complementary information during a fuel consumption oriented optimization allows both a large flexibility and the guarantee that the strategy will lower the fuel consumption.

3.7 Optimizer

The optimizer gets data regarding each mode for each segment and compares them in order to create an optimum (that could be local or absolute) in the fuel consumption. This could be done directly by dealing with estimates of the fuel consumption (explained in section 3.5), or indirectly by optimizing priorities (see 3.6).

41 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.7. OPTIMIZER

The optimizers developed during this master thesis do not aim at being integrated into the car’s electronics. The main reason is that they represent pilot algorithms in the optimization of the drive strategy that are used to observe and analyse the results of an absolute, resp. local optimum in the fuel consumption. Two different optimizers have been implemented, both of them aiming at reducing the fuel consumption:

• a global optimizer: most advanced and heaviest algorithm of the two, the global optimizer considers all possible combinations of driving modes for each segment and determines the drive strategy which gives the smallest fuel consumption. • a simplified optimizer: this faster optimizer improves a basic strategy in terms of fuel consumption up to a certain point, without giving any guarantee that the strategy will be an absolute optimum.

The optimization of a hybrid drive strategy relies on two types of data. The first type is related to the vehicle consumption. A vector summarizing the consumption in each different drive modes (HEV, EV and CHARGE) for the succession of segments is created for both the electrical and fuel consumption. The second type of data pertains to the priority of the different drive modes for the power profiles. The example of Low Emission Zones shows the importance of having priorities. This is the type of considerations that are carried by the priority vectors. For each mode a vector is generated. So far, each element of the vector corresponds to a driving category (offline priorities) but it is not excluded to change it and have for instance one element for each section (online priority), just like the consumption vectors. The priority coefficients are natural integers between 1 and 5. The higher they are, the higher the level of priority is. The choice has been made to build the algorithm around one goal of the predictive drive strategy out of the four mentioned earlier in subsection 2.1: the reduction of fuel consumption. The reason is that the fuel consumption is the core of a predictive drive strategy and should be handled as main direction of optimization. Other sub-objectives, such as improvement of the comfort and increase of the EV-experience can be taken care of by optimizer add-ons, like a relevant segmentation and the priority system.

3.7.1 Global Optimizer The global optimizer aims at determining the most fuel efficient strategy among all possible ones. This global optimizer can serve as base when developing a simpler optimizer that will not necessarily give the most fuel efficient strategy. The structure of the global optimizer as presented on Figure 3.11 shows three types of input and two output results. The algorithm is applied on an n-segment route. The inputs are:

• Consumptions: for each one of the n segments of the segmented route, the fuel (vectors * VKM Verbrauch) and electrical consumptions (vectors * ELM Verbrauch) are given in [kWh] for each of the three available modes (CHARGE, EV and HEV). • Segmentation: the category to which each of the n segments belongs is given in the integer vector Segmenttyp as an int between 1 and p, where p is the total amount of categories. Note that this illustrates a case of segmentation with external reference categories. • Priorities: one priority vector is created for each driving mode (Priorisierung EV, Priorisierung HEV and Priorisierung Charge). For the approach described on Figure 3.11, each priority vector is created offline and gives the priority of the corresponding driving mode for each of the p categories as a number between 1 and 5. 1 is the lowest priority level and 5 is the highest priority level (corresponding mode compulsory). This offline priority system is passively used, meaning that it is only integrated to a global score calculation, except for the value 5 (compulsory mode) which benefits from an extra check during the optimization process.

The outputs are:

• Strategy: integer vector of n elements giving for each of the n segments the best driving mode. 1 stands for HEV, 2 for EV and 3 for CHARGE.

42 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.7. OPTIMIZER

Figure 3.11: Structure of the global optimizer.

• SOC Gradient: float vector of n elements giving the successive levels of battery charge in [kWh] after each segment.

From the consumption, segmentation and priority inputs, the strategy and battery state of charge will be computed. The approach chosen for this global optimizer is a tree structure. As shown on Figure 3.12, each h n+1 i driving mode is tested on each segment. One node k ∈ 1; N = Pn 3j = 3 −1 represents a fictive point j=0 2 at the junction of the segments i and i + 1. It has one child for each drive mode and contains informations about the strategy taken to reach this point:

• current SOC value SOCk • priority score Priok: the higher is it, the better the priority system has been respected • strategy used to reach this point Stratk, as a vector of i integers (1 stands for HEV, 2 for EV and 3 for CHARGE) • fuel consumption in [kWh] to reach this point fk Considering node k calculated after the segment number i, the following explanations describe the process to create the 3 next nodes p, p + 1 and p + 2 resulting from the three driving modes on segment i + 1. As explained previously, the node k is characterised by four elements: SOCk, Priok, Stratk and fk. The first child to be computed is the HEV mode. The resulting node, numbered p in this example, will be calculated by assuming that the segment i + 1 is driven in HEV mode. Consequently, the parameters characterising this new node are:

• current SOC value: SOCp = SOCk + HEV ELM Verbrauch (i + 1) • priority score Priop = Priok + Priorisierung HEV (Segmenttyp (i + 1)) • strategy used to reach this point Stratp = [Stratk 1] • fuel consumption in [kWh] to reach this point fp = fk − HEV VKM Verbrauch (i + 1) The same logic is used for the EV node p + 1 and the CHARGE node p + 2. Extra verifications are performed at each step:

• check if the new SOC value lies within acceptable boundaries, usually between 20% and 80% of the battery capacity. Reasons for a restricted SOC range, such as the battery ageing and the life cycle, have been given in subsubsection 1.1.3.2. Consequently, if the value is greater than 80% of the battery capacity, it

43 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.7. OPTIMIZER

Figure 3.12: Diagram of the algorithm for the global optimizer.

is automatically brought back to this limit. If the SOC is below 20% of the battery capacity, the strategy is considered as irrelevant and no further child will stem from this node. • check if the most recent priority value has a special value. In this algorithm, it has been decided that a priority level of 5 would represent a compulsory mode. Consequently, if one mode has a priority of 5, the sons corresponding to the other drive modes will not be computed.

In the end, the last 3n nodes (in case all of them exist, cf extra verifications above) do not possess any child as they represent the result of the last segment n. They are therefore called leaves. Each leaf represents one drive strategy with its main consumption, priority and mode combination characteristics. In a final step the algorithm compares the fuel consumptions of each leaf and selects the three lowest ones. Note that it is easy to select the highest priority score in this step instead of the lowest fuel consumption, thus tuning a fuel optimizing algorithm into a priority optimizing one. The complexity of the global optimizer is exponential (proportional to 3n) and therefore quite costly in terms of time and processor load. As explained in the subsection 4.3.1 about the results of the Global Optimizer, a maximum of n = 20 segments have been optimized by the global optimizer.

3.7.2 Simplified Optimizer To answer the need of a less complex optimizing algorithm, a simplified optimizer has been programmed. The spirit of this optimizer deliberately does not aim at finding the most fuel efficient drive strategy, but rather to give one good drive strategy in terms of fuel consumption. The idea is then to compare the results of such a simplified optimizer with the global optimizer in order to assess the distance between a «good» drive strategy and the most fuel efficient ones computed by the global optimizer. The structure of the simplified optimizer is given on Figure 3.13. The algorithm is applied on an n-segment route. The inputs are:

• Basic strategy: a simple strategy, as an int vector basicStrat of n-elements (1=HEV, 2=EV and 3=CHARGE, that could be applied on the segmented route. The only requirement that this basic strategy has to meet is that it must be realizable, i.e the battery state of charge must remain between the bound- aries of 80% and 20% of the battery capacity. A basic strategy could be exclusively composed of HEV and CHARGE modes for instance, as both modes either hold or increase the battery charge. • Consumptions: for each one of the n segments of the segmented route, the electrical consumptions (vectors * ELM Verbrauch) are given in [kWh] for each of the three available modes (CHARGE, EV and HEV).

44 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.7. OPTIMIZER

Figure 3.13: Structure of the simplified optimizer.

• Segmentation: the category to which each of the n segments belongs is given in the integer vector Segmenttyp as an int between 1 and p, where p is the total amount of categories. Note that this illustrates a case of segmentation with external reference categories. The outputs are:

• Strategy: integer vector of n elements giving for each of the n segments an accurate driving mode. 1 stands for HEV, 2 for EV and 3 for CHARGE. • SOC Gradient: float vector of n elements giving the successive levels of battery charge in [kWh] after each segment. This simplified optimizer is based on the improvement of a given basic drive strategy towards a reduction of the fuel consumption. The improvement process follows the simple rule that the EV mode is more fuel efficient than the HEV mode, which in its turn has a smaller fuel consumption than the CHARGE mode. Consequently, reducing the fuel consumption here consists in an upgrade of the drive modes given by the basic strategy everytime it is physically possible. Upgrade means here changing a CHARGE mode into an HEV mode, and and HEV mode into an EV mode. Physically possible means provided that the resulting SOC remains between the 80%-20% values. The pseudo-code of the simplified algorithm is given in Algorithm1. The complexity of the simplified optimizer fit a quadratic trend to the number of segments n, as the improving strategy vector is scanned once to spot the drive modes to be upgraded, and the whole SOC gradient is computed and checked for each change.

45 CHAPTER 3. TOWARDS THE OPTIMAL PDS: METHODS 3.7. OPTIMIZER

Algorithm 1: Improvement of a basic drive strategy towards fuel efficiency. A basic strategy basicStrat is given for the n-segment route R for each segment i of the basic drive strategy basicStrat do if basicStrat(i) is CHARGE then basicStrat(i) changed to HEV if the freshly modified strategy basicStrat does not respect the SOC boundaries then basicStrat(i) changed back to CHARGE end if else if basicStrat(i) is HEV then basicStrat(i) changed to EV if the freshly modified strategy basicStrat does not respect the SOC boundaries then basicStrat(i) changed back to HEV end if end if end if end for

46 Chapter 4

Towards the optimal PDS: results

4.1 Segmentation

The segmentation will either compare the route’s segments internally or with external reference categories. Examples of categories are given in the following list:

1. Regenerative braking: more regeneration than electrical consumption

2. Super Low: Power around 0 kW

3. Low: Low power, electrical drive in HEV mode possible

4. Mid 1: no high power, pure electrical drive possible

5. Mid 2: average power but high variance

6. Mid 3: small share of high power, electrical drive mostly possible

7. High: electrical drive possible only in EV mode

8. Super High: electrical drive and charge hardly possible

9. Boost: high share of high power, frequent E-boost phases

Such a listing is an example of categories that can be used when working with external reference categories, as in the statistical parametric comparison studied here. However, this 9 category-system is not used in this work. Category numbers appearing in the following graphics are either arbitrary, as for the Kolmogorov-Smirnov function where categories are internally used, or refer to another type of category-system, like the dual city- country road system developed for the statistical parametric comparison.

Kolmogorov-Smirnov test function The Kolmogorov-Smirnov test function has been iteratively applied on the wheel power of some itineraries. The overall process is summarized in Algorithm2. Basically, the itinerary was split into same length sections, 500 meters for instance. The wheel power measurement of the first section was set as reference section. Then the 2nd section would be compared through the kstest2.m function to this reference section. If the null hypothesis of the Kolmogorov-Smirnov test was not rejected, i.e if the two signals are interpreted as stemming from the same continuous distribution, then these two sections would be considered as belonging to the same road/environment type and the third section would be compared to the reference one, etc. However, if the two densities are too different from each other, then it would be judged that a new type of environment has been found. The position of this transition would be stored and the reference profile would be reset to this new spectrum. At the end, it might be possible to have several

47 CHAPTER 4. TOWARDS THE OPTIMAL PDS: RESULTS 4.1. SEGMENTATION

Algorithm 2: Segmentation algorithm using the Kolmogorov-Smirnov test. The route R has been segmented in n 500m-long segments The first segment R(1) is saved as reference segment RS i=2 the list Bounds contains the position of the borders between two different segments while i <= n do if R(i) and RS are not similar then Bounds = [Bounds i] (i is added to the list Bounds) RS = R(i) (the new reference segment is R(i) end if i=i+1 end while separated parts corresponding to the same environment. As a result, a final comparison of all the reference profiles identified during the process would be performed. Figure 4.1 shows the result of the use of the Kolmogorov-Smirnov test algorithm on a measurement set.

Figure 4.1: Segmentation result shown on the speed [km/h] vs distance [km] measurements for an itinerary between Ingolstadt and Munich using the Kolmogorov-Smirnov test function with α = 0.05 and an initial 1000m splitting.

The comparison process has been applied to real measurements to assess the reliability of the method. The route has been measured on the 30/09/2013 from outside Munich to the city center. With an overall length of 30.029 km, it consists of a first country road part up to 15 km approximately, and then a city driving part. Although power data have been used to perform the segmentation, the graphic shows the speed as function of the distance for it is easier to understand and assess the accuracy of the model when considering speed. The first part of this itinerary has been recorded on a country road, as shown by the speeds around 80 km/h until 15 km. The frequent stops (speed decreasing to 0 km/h) denote a zone of traffic lights near the entrance of

48 CHAPTER 4. TOWARDS THE OPTIMAL PDS: RESULTS 4.1. SEGMENTATION

Munich. The rest of the data set is city driving. Notice that the close and fast oscillating speeds right after 20 km are a typical sign of congested city traffic. The numbers appearing in a green square at the bottom of the figure represent the section numbering. The section number 5 for instance is larger than the other as it includes several 1000 m long segments which have been identified as belonging to the same distribution. The numbering on the top, with grey background, gives a local classification of the sections. As mentioned previously, it is possible that two non-adjacent sections share the same distribution. Consequently, a final comparison between all the reference sections has been performed, leading to the creation of categories. For instance, sections 3 and 5 belong to the same distribution and have been gathered under the category 3. This category system is strictly local and has not been generalized with a common scale to all measurements. The results seem at a first glance acceptable. A new section has been detected after entering the city, and some significant speed reductions in the beginning have led to new limits. However, some changes are not physically meaningful. It has been wished that the first part would constitute one single section until the entrance in Munich around km 20. This is not the case, and furthermore, other noticeable speed variations have been skipped by the algorithm. The main explanation for these misinterpretations is the very high level of dependency related to the segment length and location. By simply shifting the comparison mask of a few hundred of meters, the results would drastically change. This is a recurring problem when working with energy distributions as the shape of the diagram would considerably be changed by simply adding a set of new values. To conclude, four issues were identified while analysing the results of this algorithm:

• The Kolmogorov-Smirnov test function is based on cumulative distributions. The algorithm has used power values to perform the comparison. However, the first aim was to compare energy distributions, which is different from, and not even proportional to, power cumulative distribution. • The kstest2.m function needs an alpha value, which determines the sensibility of the comparison. Results show that this sensibility factor would need to be adjusted to the section length and also to the type of environment, which is not possible as the environment is exactly what the algorithms aims to determine. • The kstest2.m function is not customizable. Should anyone want to add some comparison levels in the algorithms, the whole function would have to be written again. • Before an onboard implantation of the algorithm, a translation into ASCET (block based software used at Audi, similar to Simulink) is required. If a reprogramming is not impossible, the function kstest2.m is not the easiest to code in blocks.

Based on these observations, the choice has been made to orient the comparison algorithm towards the following functionalities:

• Easy to implement in block-based codes (such as Simulink or ASCET). • Easy to customize and adapt to constantly changing specifications. • Still based on an energy vs power histogram comparison.

Statistical parametric comparison As explained in subsection 3.4.2, 63 main parameters and some algebraic combinations between them have been computed for the 152 sections of 500 meter length belonging either to a city or a country road driving environment. Following the process detailed on Figure 3.9, the most accurate parameter that was found is the multiplication between the standard deviation of the energy distribution σ and the power P + . This parameter is Ei≥40%×Emax called Y in the following considerations, and Y = σ × P + . Ei≥40%×Emax The comparison based on the Y parameter is determined for the two categories city driving and country road. The interval for each category are built so that they contain at least 82% of the sections. Before handling the overlapping, the original intervals are Icity,0 = [0.58; 1.93] for the city driving category and Icountry,0 = [0.41; 5.97] for the country road. The original intervals are graphically represented on Figure 4.2(a). Icity,0 contains the Y values of 31 city sections out of 37 (83.7%), whereas Icountry,0 contains the values of 95 Y values out of 115

49 CHAPTER 4. TOWARDS THE OPTIMAL PDS: RESULTS 4.1. SEGMENTATION

country road sections (82.6%). The initial percentage requirement of 82% is respected, but the overlap still needs to be treated. It is immediately noticeable that Icity,0 and Icountry,0 overlap on the segment Icity-country = Icity,0 ∩Icountry,0 = [0.58; 1.93]. This step is the one where the optimization is performed: the aim is to find the parameter for which the overlap contains as few sections as possible. In this case, the overlap interval Icity-country contains Y values of 31 city sections (83.7%) and Y values of 12 country road sections (10.4%). The city category has more sections in this overlap than the country road category. As a result, the city category «wins» the overlap and can therefore keep it in its interval Icity,1 = Icity,0, whereas the overlap must be removed from the country road category Icountry,1 = Icountry,0 − Icity-country. In a nutshell, Icity,1 = [0.58; 1.93] containing 31 Y values (83.7%) for the city driving category and Icountry,1 = [0.41; 0.58[ ∪ ]1.93; 5.97] containing 83 Y values (72.2%) for the country road, as shown on Figure 4.2(b). As only two categories are considered in this study, the intervals can be extended to broader boundaries to allow a bigger amount of recognized sections. The unused interval Iunused = Icity,1 ∪ Icountry,1 = ]−∞; 0.41[ ∪ ]5.97; +∞[ can be attributed to either the city category or the country road category. One single Y value for a city section belongs to Iunused, whereas 20 country road values belong to this very same unused interval Iunused. As a result, Icountry,1 will be expanded to Icountry,2 = Icountry,1 ∪ Iunused = ]−∞; 0.58[ ∪ ]1.93; +∞[ as shown on Figure 4.2(c). The city interval remains the same: Icity,2 = Icity,1 = [0.58; 1.93]. In the end, Icity,2 contains 83.8% (31 values) of the Y parameters calculated for the 37 city driving sections and Icountry,2 contains 89.6% (103 values) of the Y parameters calculated for the 115 country road sections.

Based on the process shown in Figure 4.3, any route can technically be segmented into city or country road sections. After a segmentation of the route in equally long segments (500 m in this case), the energy vs power distributions are computed for each section and the Y parameter can be calculated from these data. Finally the categorization of the segments can be performed by comparing the Y value of each segment with the category reference intervals mentioned previously. The pseudocode of the comparison algorithm is given below.

Algorithm 3: Comparison of the Y-values with the reference intervals for each segment. The route R has been segmented in n 500m-long segments for each segment i of the route R do if Y-value of segment i in [0.58; 1.93] then the segment i belongs to the city driving category else the segment i belongs to the country road category end if end for

As for the Kolmogorov-Smirnov comparison algorithm, the comparison process is then applied to real mea- surements to assess the reliability of the method. The route is the same as previously. The use of the simple statistical parametric comparison of the Y-value gives a route segmentation as shown on Figure 4.4. The speed profile in km/h is shown on Figure 4.4(a), as function of the length in km. Each blue dot limits a 500 m section, and all sections are numbered at the bottom of the graphic in a green box. The identified categories are shown by numbers in light gray background. 2 corresponds to city driving and 3 corresponds to country road. The red dashed vertical lines limit regions of consecutive segments belonging to the same category. The power profile in kW as function of the length in km is given on on Figure 4.4(b), with the same legend as before regarding the blue points and dashed lines. It can be noticed that category 3 (country road) often appear on the first part of the route, whereas the city driving category prevails on the second part. The overall trend is therefore accurate. However, some segments are not properly identified, for both categories. Especially when driving through Munich, some 500 m long segments have been identified as country road segments. Making the most of other mathematical tools, it is possible to smooth the category variation by applying a low-pass filter that would prevent single categories when

50 CHAPTER 4. TOWARDS THE OPTIMAL PDS: RESULTS 4.1. SEGMENTATION

(a)

(b)

(c)

Figure 4.2: Interval determining process and share of segments of each kind included in the intervals: (a) original overlapping intervals, (b) disjoint intervals after overlapping treatment, (c) interval boundaries extension. appearing in minority among other ones. The result of a filter preventing «ABBA» configurations1 can be seen on Figure 4.5. If the results appear to be closer to reality, this example of the effect of a low-pass filters may not give the expected information. It is indeed interesting to keep the original segmentation in order to observe subcategories. The categorization does not only rely on road types, but can also be linked to power demand. If a city driving segment behaves like a typical country road one, it is then wiser to keep this information and adapt the hybrid configuration accordingly. The limit of this algorithm is that it highly depends on the original segmentation. In this application, segments of 500 m have been considered because the calculation of the reference Y-parameter intervals have been performed with the same length. A segmentation of the very same measurement data as already presented on Figure 4.4 has been done with a 1000 m- (see Figure 4.6(a)) and a 2000 m-segmentation (Figure 4.6(b)). Figure 4.6(a) shows the segmentation with an initial step of 1000 m. If the general trend is once again respected, new sections appear and other do not exist anymore. This step seems to favour the city driving category which is accurately identified. However, another interesting observation can be made when considering a 2000 m step as shown on Figure 4.6(b). By comparing the Y-values of 2000 m segments with the reference category intervals, the category match loses significantly in accuracy. City driving areas are identified almost

1a category A segment followed by two category B segments and then one category A segment again would be transformed into an «AAAA» route

51 CHAPTER 4. TOWARDS THE OPTIMAL PDS: RESULTS 4.1. SEGMENTATION

Figure 4.3: Statistical parametric comparison algorithm process.

(a)

(b)

Figure 4.4: Statistical parametric comparison applied on measurement data of the itinerary between Ingolstadt and Munich, 500 m-initial segmentation: (a) speed profile [km/h] vs distance [km], (b) power profile [kW] vs distance [km]. everywhere, which confirms the trend of Y-values being shifted to the city driving reference interval when the segmentation step increases. The result is an inaccurate segmentation, revealing the high dependency between

52 CHAPTER 4. TOWARDS THE OPTIMAL PDS: RESULTS 4.1. SEGMENTATION

(a)

(b)

Figure 4.5: Low-pass filter applied on the results of a statistical parametric comparison on measurement data: (a) speed profile [km/h] vs distance [km], (b) power profile [kW] vs distance [km]. the category reference intervals computed at the beginning of this statistical parametric study and the initial segmentation step. Finally, an issue can appear when considering more than just two categories. Only city driving and country roads have been implemented here because these were environments for which enough measurement data were available. But it might be interesting to take further categories into account, such as traffic jams and motor- way. More categories means more reference intervals, hence a more complicated computation process to define separate intervals and most accurate parameters to consider.

53 CHAPTER 4. TOWARDS THE OPTIMAL PDS: RESULTS 4.2. FUEL AND ELECTRICAL CONSUMPTIONS

(a)

(b)

Figure 4.6: Statistical parametric comparison applied on measurement data, speed profile [km/h] vs distance [km]: (a) 1000 m initial segmentation, (b) 2000 m initial segmentation.

4.2 Fuel and electrical consumptions

The fuel and electrical consumption topic is central as it provides the data set that will be used in the optimiza- tion. Consequently, the calculation must be precise in order to give the best consumption estimates possible. The following results deal with the power ranges to be considered when computing the fuel and electrical consumptions from the energy vs power distributions, for the EV and HEV modes.

4.2.1 Index i0 and i1

The first stage is to determine the index i0 and i1 for the drive modes EV and HEV, and for both consumption types.

EV-mode

• electrical consumption: the regenerative braking will be taken into account when calculating the electrical consumption. Consequently, the calculation will start as of the first bar, for which the braking energy can theoretically be at least partly recuperated: i0 = 1. As stated in Table 3.1, the electrical machine power is 40 kW. Due to the losses in the drivetrain between the electrical machine and the wheels, the maximal power at wheels that can be supported by the electrical machine is likely to be around 35 kW. For this reason, the electrical consumption in EV mode - which aims at using the most of the electrical machine performance - will not consider any values greater than the closest discretised interval boundary: i1 = 11 (36 kW). • fuel consumption: the ICE starts when the electrical machine cannot provide enough torque/power. As a result, i0 = 12 (36 kW) for the fuel consumption. Moreover, the ICE is involved in the next power ranges, which requires to consider the energy bars up to the end: i1 = 14 (215 kW).

54 CHAPTER 4. TOWARDS THE OPTIMAL PDS: RESULTS 4.2. FUEL AND ELECTRICAL CONSUMPTIONS

HEV-mode • electrical consumption: the regenerative braking will be taken into account when calculating the electrical consumption. Consequently, the calculation will start as of the first bar, for which the braking energy can theoretically be at least partly recuperated: i0 = 1. The HEV mode possesses a limit around 20 kW above which the ICE will be switched on. Converting this limit to energy at wheels gives approximately 18 kW, meaning that energy bars will be considered up to the closest power range, i.e i1 = 7 (20 kW). • fuel consumption: the ICE starts when the electrical machine has reached the HEV threshold. As a result, i0 = 8 (20 kW) for the fuel consumption. Moreover, the ICE is involved in the next power ranges, which requires to consider the energy bars up to the end: i1 = 14 (215 kW). Note that these indexes can always be easily adjusted if needed thanks to the configuration of the energy vs power distribution. Table 4.1 summarizes the index i0 and i1 for the two consumption types and the two drive modes, and Figure 4.7 shows a graphical representation of it.

Table 4.1: Index i0 and i1 to be used for calculation of the electrical and fuel consumptions, for EV and HEV modes. EV mode HEV mode electrical fuel electrical fuel i0 1 12 1 8 Power -100 kW 36 kW -100 kW 20 kW i1 11 14 7 14 Power 36 kW 215 kW 20 kW 215 kW

In theory, the battery is also involved during the E-boost (last bar). As a result, this power range should be included in the computation of the electrical consumption. However, the exact contribution of the electrical machine in this feature has not been determined precisely and is therefore not considered here.

55 CHAPTER 4. TOWARDS THE OPTIMAL PDS: RESULTS 4.2. FUEL AND ELECTRICAL CONSUMPTIONS

e: electrical consumption l: fuel consumption EV

(a) (b) HEV

(c) (d)

Figure 4.7: Values to be considered for the consumption computation (red bars): in EV mode, electrical (a) and fuel (b) - in HEV mode, electrical (c) and fuel (d).

4.2.2 Conversion values As explained in section 3.5, one vector of conversion values is defined for the electrical consumption and one is defined for the fuel consumption. If the values have not been calculated yet, the method to compute them has been determined: 1. For each of the 14 ranges of power in the energy vs power distribution calculate the combined efficiency of the electrical drivetrain (battery, electrical machine, gearbox) and the combined efficiency of the traditional drivetrain (engine, gearbox). Note that the efficiency may vary considerably between the points, even though they belong to the same narrow power range. In this case, an average can be performed provided that its results are confirmed by a comparison with real measurements. 2. For the two first bars, which pertain to regenerative braking, determine the contribution of energy that can be regenerated with regards to the contribution of mechanical brakes. 3. For the last bar, pertaining to E-boost, determine the contribution of the electrical machine vs ICE. As final stage, it is relevant to use the 28 categorized measured segments performed at the beginning of this thesis to compare the real values of e and f with the computation in order to assess the accuracy of the

(αi)i=1..14 and (βi)i=1..14.

56 CHAPTER 4. TOWARDS THE OPTIMAL PDS: RESULTS 4.3. OPTIMAL DRIVE STRATEGY

4.3 Optimal drive strategy 4.3.1 Global optimum Three implementations The global optimizer is coded in different languages/software programs:

• Matlab: the development and improvements of the global optimizer have been performed using Matlab. One reason is that this numerical calculation language accepts almost any input data type for importation, such as the .dat files that are provided by the car’s onboard measurement unit. Another significant asset of Matlab is the possibility of displaying graphically informations such as diagrams and curves (see Figure 4.8(a)). This allows quick data checks as well as an easy-to-understand representation of the results, even though the whole process is CPU guzzling. • C: the global optimizer has been coded in C-language, as transitional step between Matlab and ASCET. The code is available in Appendix 7.2.1. The implementation in C using CodeBlocks is much faster than in Matlab, as a heavy software program is not needed and C directly uses the computer resources without going through a whole user-friendly interface. However, the representation of the results is not user-friendly and is done using a text format displayed on the console, as shown on Figure 4.8(b). • ASCET: the final form of the global optimizer was coded is ASCET, which is the main program used by Audi for function development (see Appendix 7.2.2). ASCET is a Simulink-alike tool, based on connected blocks and standardized inputs/outputs to guarantee a safe use of the functions. In addition to traditional blocks, ASCET allows the use of C-coded classes and modules as long as they meet the input/output standards of the software program. The implementation of the global optimizer on ASCET is slightly slower than in C. The experimentation interface is not as customizable as in Matlab, but relatively easy to arrange and very familiar to all engineers working in the department. In experimentation mode, vectors and parameters are shown using measurement tools such as numerical displays (see Figure 4.8(c)).

Table 4.2 gives an overall comparison between the different implementations.

Table 4.2: Comparison between the three implementations of the global optimizer.

Matlab C ASCET Computation time for a 20-segment route 3 days 30 seconds 1 minute Programming type scripts and functions classes classes/blocks Result representation graphical text numerical Language/Software maths-oriented universal widespread at Audi

Result analysis Because all previously mentioned implementations are based on the very same algorithms, the following analysis will use Matlab graphical representations to make is as clear as possible. The implantation in Matlab shows the three most fuel efficient drive strategies as battery state of charge [kWh] gradient along the route. Figure 4.9 shows the graphical representation of the results for a route coded as 2-3-4-3-2-3-4-3-2. The number 2 represents a segment measured in a city driving environment, 3 stands for a country road and 4 for motorway. For each of these three measured segments, the electrical and fuel consumptions of EV and HEV modes have been computed using a Simulink model of an Audi PHEV e-tron project2. The CHARGE mode characteristics have not been defined which prevents any estimates for this drive mode. Consequently, the CHARGE mode is excluded from the following simulations. These measured segments have been concatenated to create an artificial route simulating a travel between two cities with a stop in an intermediate city, all connected by country roads and motorways.

2 The method explained in section 3.5 has not been implemented yet as the conversion parameters (αi) and (βi) are still undetermined.

57 CHAPTER 4. TOWARDS THE OPTIMAL PDS: RESULTS 4.3. OPTIMAL DRIVE STRATEGY

(a)

(b)

(c)

Figure 4.8: Three different implementations of the global optimizer and the representation of the results: (a) Matlab, (b) C, (c) ASCET.

The top SOC gradient on Figure 4.9 represents the most fuel efficient drive strategy, while the middle respectively bottom ones represent the second respectively third best strategies. On each gradient, the segment identification (2, 3 or 4) appears in grey boxes on top and a vertical dashed line symbolizes the limit between two segments. The battery state of charge is shown along the y-axis in [kWh] and the route is displayed along

58 CHAPTER 4. TOWARDS THE OPTIMAL PDS: RESULTS 4.3. OPTIMAL DRIVE STRATEGY

(a)

(b)

(c)

Figure 4.9: Global optima for the route 2-3-4-3-2-3-4-3-2 as displayed by Matlab. SOC profiles [kWh] vs segment sequence: (a) 1st most fuel efficient drive strategy, (b) 2nd most fuel efficient drive strategy, (c) 3rd most fuel efficient drive strategy. the x-axis. All SOC gradients start at maximum battery charge, which is 80% of the total battery capacity (8.54 kWh approximately). As explained in section 3.7, one condition about SOC values is to remain between 20% and 80% of the total battery capacity. The SOC gradient segment color represent the driving mode suggested on this segment. Blue stands for HEV mode, green for EV mode and red for CHARGE mode. Note that no red segment can be seen as the CHARGE mode is not implemented in the input consumption values and can therefore not be integrates in the optimization. At the bottom of each SOC gradient, important information about the drive strategy are display in a grey background banner, also shown on Table 4.4:

• the drive strategy: a string of H for HEV mode, E for EV mode and C for CHARGE mode (when available) represents the drive strategy corresponding to the SOC gradient displayed above it. • Score: the first number shows the priority score of the corresponding drive strategy, i.e the sum of the priorities given to the chosen drive modes for each segment. The second number is the maximum priority score that can be reached if all preferred drive mode were chosen for each segment. As explained previously, the optimization does not take into account the priorities as only the fuel consumption is optimized. However, this can be changed easily in the algorithm. Table 4.3 gives the priority points for each drive mode and each of the three categories (2, 3 and 4). • KSV: stands for Kraftstoffverbrauch which means fuel consumption. The value following KSV represents

59 CHAPTER 4. TOWARDS THE OPTIMAL PDS: RESULTS 4.3. OPTIMAL DRIVE STRATEGY

the overall fuel consumption in [kWh] on the given route with the chosen drive strategy. • ELM: gives the electrical consumption in [kWh] (energy taken from the battery) on the given route with the chosen drive strategy.

Table 4.3: Priority points for each drive mode and driving categories 2, 3 and 4.

2 3 4 HEV 0 2 2 EV 2 1 0 CHARGE 1 1 1

Table 4.4: Summary of the three best drive strategies for the route 2-3-4-3-2-3-4-3-2.

1st Best 2nd Best 3rd Best Drive strategy H-H-E-H-H-H-H-E-H H-E-H-H-H-H-E-H-H E-H-H-H-E-H-E-H-E Priority score 9/18 9/18 16/18 Fuel consumption [kWh] 205.5597 205.5597 205.9506 E-consumption [kWh] 6.3585 6.3585 5.8134

The main details of each optimal drive strategy are given again in Table 4.4. It is here interesting to note that the two first strategies only differ from two mode switches. In the best drive strategy, the segment number 2 (of type 3) is driven in HEV mode and the segment number 8 (type 3) is driven in EV mode. It is the opposite in the second best drive strategy. A similar inversion exists for two segments of type 4 (located in positions 3 and 7). Consequently, both drive strategy have the same priority score, fuel consumption and electrical consumption. However, one of them appears safer in terms of electrical drive. The first one keeps a high SOC level until the 7th segment, i.e almost until the end. This strategy is interesting as it allows some unexpected deviations in real life from the predictions without implying a complete change in the drive strategy. On the contrary, if the second segment is run in EV mode as suggested by the second drive strategy, an electrical consumption slightly higher than expected might make the SOC level reach the limit value of 20%, hence activating a battery critical SOC emergency mode which would charge the battery. For this reason, keeping a high SOC level as much as possible brings an extra safety regarding the respect of the predicted drive modes. Note that such a situation where the two best drive strategies have identical characteristics technically never happens, as it is normally seldom that one segment is identically reproduced in a route. Moreover, the three strategies are quite close in terms of consumption, which is also due to the fact that the 9-segment route has been artificially computed by concatenating 3 reference segments. Consequently, each segment appear several times and thus allowing several combinations of drive modes leading to similar results. In the several optimizations run with different routes, recurrent associations between modes and categories have not appeared frequently enough, for a rule to be set. In this previous example, the two best drive strategies run all city segments (type 2) in HEV mode, whereas the third strategy applies an EV mode on it. It would have been interesting to note some mode-category associations coming up often, in order to include it in an offline prioritization.

Qualifying the algorithm As this tree-based algorithm actually relies on the search of the best drive strategy - meaning the problem’s global optimum - it is quite costly in terms of time and processor load. The use of such a data structure allows to use recursive algorithms. However, the exponential complexity makes it difficult to apply the algorithm on itineraries with several sections identified. For instance, computing a 12 section route would take approximately 30 seconds on Matlab (the implementation in C is faster as 19 sections would be computed in the same time) and this time would triple for each extra section.

60 CHAPTER 4. TOWARDS THE OPTIMAL PDS: RESULTS 4.3. OPTIMAL DRIVE STRATEGY

Once the tree has been created, it is easy to visit each leaf (last nodes) and to compare the priority scores or the fuel consumptions in order to find the best strategies. The biggest advantage but also inconvenient of this algorithm is that it embraces all the possible solutions. Consequently, the results displayed are for sure the optima with regard to the strategy. However, calculating the 3n possibilities also takes a certain time. An idea to keep the same algorithm but use it with more flexibility is to work on the minimum segment lengths defined for the segmentation part (see section 3.4). Instead of revising and re-establishing the drive mode every kilometre for instance, it would be much faster to perform the optimization on 10 to 50 km long segments. Consequently, a 12 section itinerary would measure between 120 and 600 km. After such a distance, it would be interesting to recharge the battery of the Plug-In Hybrid vehicle. The second part of the itinerary could then be calculated separately. Instead of optimizing the drive strategy for 24 sections (200 days of calculation on a regular computer) for instance, the calculation would be split into 2 optimizations of 12 section each (2 x 30 seconds). This optimizer has then been coded in C and integrated into the ASCET model of the Long Range Prediction Function. A pleasant surprise was to realize that the performance increases when running a C-code compared to Matlab. As a result, optimizing a 20-section itinerary would only take a little bit more than one minute with this new version. The limitation regarding the number of sections to be optimized has become slightly less limiting for the application, but still represents an issue for a possible future onboard implementation. The question of the use of fuel consumption data has been raised after the implantation of this optimizer. This information is used as optimization target. In other words, the total fuel consumption is the variable that will be minimized during the process. However, it might as well be interesting to consider a system of priority. If all section types are given a certain priority for each drive mode, it will then be possible to optimize and assess the overall priority level of the strategy. Reasons for considering priorities instead of fuel consumption are:

• Using priorities may reduce the complexity of the algorithm. Instead of trying all the available strategies for each and every sections, the algorithm could optimize all the sections belonging to a same type at a time. For instance, it could give all traffic sections an electric mode. • An optimization directly based on the consumption requires predictive consumption data. The optimizer needs to know how much fuel will be consumed on each section. If computing the electrical consumption is relatively easy, it remains quite difficult to determine the fuel consumption. This is mainly due to the various operating points for which the ICE can be used, as well as the driver’s behavior, . . . Moreover, the accuracy of the fuel consumption estimates must be very high to ensure that the strategy computed is the real optimal one. • As mentioned previously, a priority system also allows to take into account other parameters than fuel consumption, such as EV-experience and imperatives pertaining to road regulations (ZEZ and LEZ).

4.3.2 Enhanced basic strategy The simplified optimizer has been implemented in Matlab in order to assess the distance between an enhanced basic drive strategy and the optimal drive strategies provided by the global optimizer. The graphical features of Matlab allow a visual comparison between several SOC gradients. The basic strategy used as base for the optimization is an all HEV strategy, except for the city segments (type 2) where the EV mode is requested.

Result analysis The implementation on Matlab shows the three most fuel efficient drive strategies from the global optimizer and the enhanced basic strategy from the simplified optimizer. Figure 4.10 shows the graphical representation of the results for the route taken as example with the global optimizer previously. The same comments apply to the graphical layout of the Matlab representation. Note that the priority score has been computed externally for the simplified optimizer for comparison purpose, though priorities are not taken into account by the algorithm. The main details of each drive strategy are given again in Table 4.5, EBDS standing for Enhanced Basic Drive Strategy. It is immediately noticeable that the 3rd most fuel efficient drive strategy (shown on Figure 4.10(c))

61 CHAPTER 4. TOWARDS THE OPTIMAL PDS: RESULTS 4.3. OPTIMAL DRIVE STRATEGY

(a)

(b)

(c)

(d)

Figure 4.10: Global optima and enhanced basic drive strategy for the route 2-3-4-3-2-3-4-3-2 as displayed by Matlab. SOC profiles [kWh] vs segment sequence: (a) 1st most fuel efficient drive strategy, (b) 2nd most fuel efficient drive strategy, (c) 3rd most fuel efficient drive strategy, (d) enhanced drive strategy provided by the simplified optimizer.

Table 4.5: Summary of the four drive strategies for the route 2-3-4-3-2-3-4-3-2.

1st Best 2nd Best 3rd Best EBDS Drive strategy HHEHHHHEH HEHHHHEHH EHHHEHEHE EHHHEHEHE Priority score 9/18 9/18 16/18 16/18 Fuel consumption [kWh] 205.5597 205.5597 205.9506 205.9506 E-consumption [kWh] 6.3585 6.3585 5.8134 5.8134 CPU-time [s] 2.18 0.016

and the enhanced basic drive strategy (Figure 4.10(d)) are identical. This example has been chosen to prove that an enhanced drive strategy can be very close to an optimal one, while keeping a reasonable computation time and CPU demand. The simplified optimizer only performed a very limited number of changes. The basic strategy given as input was very close to the final resulting drive strategy as shown on Table 4.6. Only the 7th segment is different, as it was originally set as HEV. The single segment that could be changed by the simplified optimizer was this 7th one (type 4 motorway). Note that the other type 4 segment located in position 3 could not be changed from

62 CHAPTER 4. TOWARDS THE OPTIMAL PDS: RESULTS 4.4. RESULTING FUEL SAVINGS

HEV to EV as this would have lowered the SOC at the end of the 5th segment under the 20% limit. Table 4.6: Basic and enhanced basic drive strategies for the route 2-3-4-3-2-3-4-3-2.

1 2 3 4 5 6 7 8 9 Basic drive strategy E H H H E H H H E Enhanced basic drive strategy E H H H E H E H E

Qualifying the algorithm Although the results shown here are quite convincing, this is mainly due to the relevance of the basic drive strategy, which determined 90% of the modes of the enhanced basic drive strategy. However, the choice made for the basic drive strategy was risky too, as setting all city categories to EV modes might not always be an acceptable strategy. Consequently, the basic strategy is a topic that needs to be further studied in order to come to simple rules to establish it. A direction of development could be based on the potential of electrical energy generation of each segment. When for instance a city segment is located after country road segments where a standard HEV mode usually generates more electrical energy than what is consumed, then that city segment is likely to find enough resources to be run on EV mode. When comparing the complexity and computation time of the simplified optimizer and the global optimizer, it appears that a simple algorithm will be easier and more flexible to be integrated onboard on Plug-In Hybrid cars. Many other options for a simple optimization are to be explored and their results need to be compared with the global optima is order to choose the most relevant one for the application.

4.4 Resulting fuel savings

As mentioned in section 2.1, one of the goals of a predictive driving strategy is the reduction of the fuel consumptions. The fuel savings resulting from a predictive drive strategy depend on the route. They are mostly done in urban environments where a relevant hybrid strategy can prevent the use of the ICE when the load points have a low efficiency. Basic fuel consumption comparisons can be performed based on the example used all along this subsection about optimization. The results are presented in Table 4.7. Because a large part of the fuel economy for hybrid cars is made in urban environments, the difference between the standard version of the Audi A6 and the modified hybrid model is rather small for this route mainly constituted of country road and motorway profiles. Table 4.7: Fuel consumptions and savings between the different drive strategies for the route 2-3-4-3-2-3-4-3-2.

1st -2nd Best 3rd Best-EBDS Std Hybrid Std Gasoline Fuel consumption [kWh] 205.5597 205.9506 227.6679 271.8280 Fuel consumption [l] 21.2603 21.3008 23.5469 28.1142 Fuel savings/Basic Hybrid [%] -9.7 -9.5 - +19.4 Fuel savings/standard vehicle [%] -24.4 -24.2 -16.3 -

Even if predictive drive strategies have been computed on other artificially-made routes, the next step is to optimize the strategy on real itineraries. The objective is to drive and measure a route in EV mode, then HEV mode and finally CHARGE mode. Then the segmentation algorithm would split it and the fuel and electrical consumptions for each of the three drive modes would be exactly known for each segment. From there on, the consumption estimates could be compared to the real measurements. Moreover, the optimizer based on real measurements would also provide a certain optimal drive strategy. This strategy could be driven on the route and the overall consumption could finally be compared to the estimates to assess the accuracy of the methods.

63 64 Chapter 5

Conclusion

5.1 Predictive drive strategy implementation

A predictive drive strategy is the setting of a combination of drive modes beforehand, in order to achieve reductions in the fuel consumption, to enhance the EV-experience, to improve the comfort and to remove the need of a manual switch of mode. In this master thesis dealing with the optimization of the predictive drive for a PHEV, the whole process including the route’s predictive power data, the route segmentation, the fuel/electrical consumptions of a seg- ment, the drive mode/drive category prioritization and the optimization of a drive mode combination has been covered. If the measurements and models focus on an Audi A6 Hybrid, the methods aim at being universal and applicable to any kind of Plug-In Hybrid vehicle. The layout of the predictive power data has been determined using predictions pertaining to the road topography, speed limits and real-time traffic information. The energy vs power distribution representation has the advantage of being time-independent, thus being a universal structure for all types of sections. This allows comparisons between different distributions used to segment the route as well as consumption computations. The segmentation is a central topic in predictive drive strategy. It is the first step performed on the predictive data and has a determining influence on the following stages. The segmentation aims at splitting a given route into segments of a certain length. Then a comparison method is used to compare their characteristics. Two types of segmentation can be used: an internal segmentation comparing the segments with each other within the route in order to determine if some are similar or not, and an external segmentation comparing the segments with reference categories such as city driving, country road or motorway. If the overall process has been determined, further studies and developments still need to be done to find out what comparison algorithm should be used. The Kolmogorov-Smirnov has the potential to give accurate and relevant results if the parameters ruling the algorithm, such as the sensibility represented by the α-value are correctly set. A statistical parametric comparison can give good results with the identification of one category among the segments. However, the study becomes significantly heavier when several categories are at stake. Consequently, the Kolmogorov-Smirnov method should be preferred here, at least until further results are obtained. The structure of the main data support, the energy vs power distribution, allows a flexible exploitation and can also be used to compute the fuel and electrical consumptions related to the corresponding segment, for different drive modes. The HEV, EV and CHARGE modes all have different configurations in terms of electrical drive limits and use of the ICE for charging the battery. The histogram structure offers the possibility to take these differences into account. The challenge is to cover the energies at the wheel into energies out of the tank or out of the battery. The mathematical approach is defined, using a conversion coefficient for each bar and then summing the converted energies in a power domain that fits the corresponding drive mode. The values of each (αi) (electrical consumption) and (βi)(fuel consumption) coefficient vectors are to be determined considering three elements: the efficiency chain from the wheels to the tank/battery, the regenerative braking parameters and the electrical contribution in the E-boost feature. Prioritizing drive modes depending on the environment would allow more flexibility and performance towards

65 CHAPTER 5. CONCLUSION 5.2. FUTURE WORK

an optimized drive strategy. The idea here is to associate a drive category to a list of the drive modes, sorted from the most to the least relevant. That way, other targets than the fuel consumption could be considered in the optimization, while offering a faster optimization. Two types of prioritizations have been identified: offline priorities, which work with predefined categories like in the comparison with external categories, and online priorities, which are calculated internally using the fuel and electrical consumptions for instance. In addition to this, priorities can be used as complementary information that do not drive the optimization but rather represent further constraints in the optimization problem. This allows to integrate possible Zero Emission Zones to the optimization, where an EV mode is absolutely mandatory. The choice of the priority system is closely related to the choice in the optimizer method. The final step in the computation of an optimal drive strategy is the calculation in itself, which is performed by an optimizer. The optimizer takes electrical and fuel consumptions, priorities and the segmented route as inputs and computes the optimal drive strategy and the corresponding SOC gradient. The choice was made to optimize the fuel consumption, but other items may be considered as mentioned previously. Two optimizers have been programmed and tested: a global optimized as pilot algorithm to determine the absolute best drive strategies, and a simplified optimizer which improves a basic strategy. The global optimizer has been coded in Matlab, C and ASCET in order to be used as reference. It does not aim at being integrated into the car’s onboard system, but serves as basis to compare enhanced strategies from simple algorithms with absolute optima. Because the global optimizer is heavy and resources-demanding, the future PHEV are likely to be equipped with simplified optimizers. The simplified optimizer developed in this master thesis improves a basic drive strategy that it gets as an input by upgrading the modes towards fuel efficiency. The quadratic complexity allows much more computation flexibility but so far no guarantee can be given regarding the distance in terms of optimality between this improved strategy and the most fuel efficient ones. Significant fuel savings are expected with the implementation of predictive drive strategies. The calculations performed in this master thesis give a gain of 10% in fuel savings compared to a traditional hybrid. However, the reduction of fuel consumption is only one of the four goals of the predictive drive strategy. To offer a broader EV- experience and respect future legislations making it mandatory to drive all electric through some environmental zones, a predictive drive strategy can manage the energy resources in order to make them available when needed. In a nutshell, predictive drive strategies will bring a new, economical, more automated and smoother way of driving while enhancing the role of Plug-In Hybrid vehicles towards greener transportation.

5.2 Future work

The future work will focus around the four axes of predictive driving strategy computation: the route segmen- tation, the fuel and electrical consumptions, the priority system and the optimizer. The energy vs power distribution is the information carrier. Based on this data format, the route will be segmented, most likely using a method similar to the Kolmogorov-Smirnov. More tests need to be performed with this function in order to understand the effect of the alpha-value and assess the accuracy of the segmentation. In addition to the function, a questioning about the segment categories is required. It is important to decide if the driving environment needs to be determined based on the speed limit for instance, or on the energy demands. Regarding the fuel and electrical consumption calculations, a thorough model of the drivetrain including the gearbox, electrical machine and ICE has to be used to determine the conversion coefficients which aims at converting the energy at wheels to energy taken from the tank or from the battery. Not only the whole efficiency chain needs to be considered, but also the parametrization of the regenerative braking and the E-boost, so that the respective contributions of the battery and the fuel can be known. The possibility of using prioritization of the drive modes for each segment or category needs to be further explored. Priorities bring an extra flexibility and considerably simplify the optimization processes. However, the type of prioritization - offline or online - has to be decided first. Then, the association drive mode to segment type will be determined through mathematical computations involving the fuel and electrical consumption, through a statistical analysis of measurements or simply with regards to the comfort and EV-experience. The optimizer that is likely to be integrated in the future Audi PHEV will probably be similar to the simplified algorithm presented in this thesis. The requirements in terms of CPU prevents an extended use of a

66 CHAPTER 5. CONCLUSION 5.2. FUTURE WORK

global optimizer. The algorithm of this simplified optimizer will be determined considering the efficiency of the computed predictive driving strategies and the time needed to obtain it. For each new pilot simplified optimizer, the results will be compared with those stemming from the global optimizer. A solid base has been developed for methods towards the optimization of a predictive drive strategy. If some aspects of it need to be finalized and decided, the possibilities have been identified and the expected results are known. The future works will be mostly based on physical and empirical approaches as this will be the final step before an onboard implementation of the system.

67 Chapter 6

Bibliography and tables

68 Bibliography

[1] Daniel Sperling and Deborah Gordon. Two Billion Cars: Driving Toward Sustainability. Oxford University Press, 2009.

[2] Yunshi Wang, Jacob Teter, and Daniel Sterling. Will china’s vehicle population grow even faster than forecasted? Access #41 pp 29-33, 2012.

[3] Hong Huo, Michael Wang, Larry Johnson, and Dongquan He. Projection of Chinese motor vehicle growth, Oil demand and CO2 emissions through 2050. ANL/ESD/06-6, Argonne National Laboratory, Energy Systems Division, 2006.

[4] Audi AG. Audi in Zahlen December 2013. Audi internal document, 2013.

[5] Jean-Paul Rodrigue, Claude Comtois, and Brian Slack. The Geography of Transport Systems - 3rd Edition. Routeledge, 2013.

[6] Official Journal of the European Communities. Council directive 91/441/EEC. 26 June 1991.

[7] Xinhua. China to adopt auto emission standard equal to euro iii in 2008. China Daily, 2004.

[8] Ministry of Environment and Forests Government of India. Press release #65196: Emission norms for Petrol & Diesel in India. 2010.

[9] Regulation no 443/2009 of the European Parliament and of the Council. 23 April 2009.

[10] European Environment Agency. CO2 emissions of car manufacturers in 2012. EEA, 2012.

[11] Transport & Environment. How Clean Are Europe’s Cars? T&E, 2013.

[12] IHS Automotive. The Advanced Automotive Battery Report. IHS, 2014.

[13] Niklas Winkler. MF2047: Lecture 1, KTH. Lecture notes, 2012.

[14] Bosch. Automotive Handbook, 8th Edition. Wiley, 2011.

[15] Per Risberg. MF2047: The Characteristics of the Otto Engine, KTH. Lecture notes, 2012.

[16] Timothy E. Lipman and Mark A. Delucchi. Emissions of nitrous oxide and methane from conventional and alternative fuel motor vehicles. Climatic Change pp 477-516, Kluwer Academic Publishers, 2002.

[17] Lars J. Pettersson. MF2047: Catalytic Automotive Pollution Control, KTH. Lecture notes, 2012.

[18] Alain HAUPAIS. Combustion dans les moteurs Diesel. Technique de l’ingenieur, 1999.

[19] Per Risberg. MF2047: The Characteristics of the Diesel Engine, KTH. Lecture notes, 2012.

[20] AutoUni. E-Traktion: Getriebe in Hybrid und E-Fahrzeugen. AutoUni, 2012.

69 BIBLIOGRAPHY BIBLIOGRAPHY

[21] Markus Kaun. Funktionsentwicklung & Applikation einer Fahrsituationserkennung für die Start-Stopp Funktion durch Nutzung vernetzter Informationen. TU Berlin - Master Thesis, 2013. [22] Boris Blasinski. Optimierung und Integration der Betriebsstrategie eines Kraftfahrzeugs mit parallelhybri- dem Antriebsstrang. Fachhochschule Ingolstadt - Diplomarbeit, 2008. [23] Michael Back. Pr¨adiktiveAntriebsreglung zum energieoptimalen Betrieb von Hybridfahrzeugen. Universit¨at Karlsruhe - PhD Dissertation, 2005. [24] Audi AG. Audi Q5 Hybrid Quattro, Selbststudienprogramm 489. Audi internal document, 2011.

[25] Audi AG Kommunikation Produkt. Audi Showcars Genf 2010. Press release, 2010. [26] Controlled Power Technologies. CPT SpeedStart System Description. Press release, 2013. [27] Audi AG. Bericht zur Wirtschaftslage Wesentliche Unternehmensdaten der AUDI AG für den Zeitraum Januar bis Dezember 2013. Audi internal document, 2014.

[28] Audi AG. Gesch¨aftsbericht 2013. Audi internal document, 2014. [29] Michael Schraut. Umgebungserfassung auf Basis lernender digitaler Karten zur vorausschauenden Kondi- tionierung von Fahrerassistenzsystemen. TU M¨unchen - PhD Dissertation, 2000. [30] BMW Group. BMW Group Innovation Day 2012: Efficient Dynamics, Media Information. Press release, 2012. [31] Kristian Billig, Manfred Klüting, Carsten Breitfeld, Stefan Juraschek, and Fritz Steinparzer. The Innovative BMW Powertrain in the All New BMW i8, BMW Group. 22nd Aachen Colloquium Automobile and Engine Technology - Conference, 2013.

[32] Jens Papajewski. Prädiktive Betriebsstrategie für Plug-In Hybridfahrzeuge mit Ottomotor, Audi AG. Audi internal document, 2013. [33] Marcus Sonntag and Oliver Sawodny. Vorausschauende Antriebsstrangregelung f¨urE-Hybrid LKW im Fernverkehr. Project, 2011. [34] Audi AG. Audi A6 technical data. Technical datasheet, 2013.

[35] T J Barlow, S Latham, I S McCrae, and P G Boulter. A reference book of driving cycles for use in the measurement of road vehicle emissions. TRL, 2012. [36] International Council on Clean Transportation. World-Harmonized Light-Duty Vehicles Test Procedure. ICCT - Policy Update, November 2013.

[37] United Nations Economic Commission for Europe Inland Transport Committee. World Forum for Harmo- nization of Vehicle Regulations 162nd session. Annotated Agenda, 10 December 2013. [38] KTH Vehicle Engineering Department. SD2222: Chapter 1 - Wheels, Tyres and Motion Resistance. Lecture notes, 2012.

[39] M. Brandstäter, G. Bauer, and W. Prestl. Acc funktionsoptimierung durch nutzung von navigations-daten. VDI-Berichte, 2003.

70 List of Figures

1.1 Schematic energetic diagram of the E-boost [22]...... 7 1.2 Schematic energetic diagram of the regenerative braking [22]...... 7 1.3 Engine map and effect of a load point increase/decrease [23]...... 8 1.4 Schematic energetic diagram of a lower load point (a) and a higher load point [22]...... 8 1.5 Schematic energetic diagram of the full electric mode [22]...... 9 1.6 Schematic energetic diagram of the normal drive [22]...... 9 1.7 Schematic diagram of a series hybrid drivetrain [23]...... 10 1.8 Schematic diagram of a parallel hybrid drivetrain [23]...... 11 1.9 Schematic diagram of a power-split hybrid drivetrain [23]...... 12 1.10 Energy density of some fuels and batteries [12]...... 13 1.11 Volvo’s Continuous Electric Drive system, Volvocars news webpage...... 14 1.12 Chart of the VW Group and its subsidiaries...... 15 1.13 Shares of Audi models in the 2013 sales [27]...... 16 1.14 Organization chart of the Powertrain Development Department at Audi, focus on the group EA-242 where the thesis has been conducted...... 16 1.15 Audi A1 e-tron...... 17 1.16 Audi A3 e-tron...... 17 1.17 Audi A6 L...... 18 1.18 Audi Sport Quattro Concept...... 18

3.1 Speed profile of the NEDC...... 27 3.2 Speed profile of the FTP-75...... 28 3.3 Speed profile of the WLTP [36]...... 28 3.4 Representation of the forces applied on a car [38]...... 29 3.5 Ratio between the equivalent mass of the rotating parts and the empty mass of the A6, as function of the total gear ratio...... 31 3.6 Energy distribution [kWh/100km] vs power [kW] for the NEDC cycle (a) and the FTP-75 cycle, for an Audi A6...... 33 3.7 Global segmentation process...... 34 3.8 Example of the three different types of intervals that can be found for a given parameter (called X) and two categories (Country road and City)...... 37 3.9 Process for statistical comparison...... 38 3.10 Example of values to be taken into account for the consumption computation (red bars): in EV mode, electrical (a) and fuel (b) - in HEV mode, electrical (c) and fuel (d)...... 40 3.11 Structure of the global optimizer...... 43 3.12 Diagram of the algorithm for the global optimizer...... 44 3.13 Structure of the simplified optimizer...... 45

71 LIST OF FIGURES LIST OF FIGURES

4.1 Segmentation result shown on the speed [km/h] vs distance [km] measurements for an itinerary between Ingolstadt and Munich using the Kolmogorov-Smirnov test function with α = 0.05 and an initial 1000m splitting...... 48 4.2 Interval determining process and share of segments of each kind included in the intervals: (a) origi- nal overlapping intervals, (b) disjoint intervals after overlapping treatment, (c) interval boundaries extension...... 51 4.3 Statistical parametric comparison algorithm process...... 52 4.4 Statistical parametric comparison applied on measurement data of the itinerary between Ingol- stadt and Munich, 500 m-initial segmentation: (a) speed profile [km/h] vs distance [km], (b) power profile [kW] vs distance [km]...... 52 4.5 Low-pass filter applied on the results of a statistical parametric comparison on measurement data: (a) speed profile [km/h] vs distance [km], (b) power profile [kW] vs distance [km]...... 53 4.6 Statistical parametric comparison applied on measurement data, speed profile [km/h] vs distance [km]: (a) 1000 m initial segmentation, (b) 2000 m initial segmentation...... 54 4.7 Values to be considered for the consumption computation (red bars): in EV mode, electrical (a) and fuel (b) - in HEV mode, electrical (c) and fuel (d)...... 56 4.8 Three different implementations of the global optimizer and the representation of the results: (a) Matlab, (b) C, (c) ASCET...... 58 4.9 Global optima for the route 2-3-4-3-2-3-4-3-2 as displayed by Matlab. SOC profiles [kWh] vs seg- ment sequence: (a) 1st most fuel efficient drive strategy, (b) 2nd most fuel efficient drive strategy, (c) 3rd most fuel efficient drive strategy...... 59 4.10 Global optima and enhanced basic drive strategy for the route 2-3-4-3-2-3-4-3-2 as displayed by Matlab. SOC profiles [kWh] vs segment sequence: (a) 1st most fuel efficient drive strategy, (b) 2nd most fuel efficient drive strategy, (c) 3rd most fuel efficient drive strategy, (d) enhanced drive strategy provided by the simplified optimizer...... 62

7.1 Overview of the Audi fleet 2013...... 78 7.2 Overview of the predictive drive strategy algorithm in ASCET...... 84 7.3 Predictive drive strategy optimizer in ASCET...... 85 7.4 Predictive drive strategy consumption unit in ASCET...... 86

72 List of Tables

1.1 European emission standards for a passenger car (g/km), Diesel...... 2 1.2 European emission standards for a passenger car (g/km), Gasoline...... 2 1.3 Distance to target for the VW pool in 2012 [10]...... 3 1.4 Hybridization degrees [20]...... 6 1.5 A1 e-tron technical information...... 17 1.6 A3 e-tron technical information...... 17 1.7 A6 L e-tron concept technical information...... 18 1.8 Sport Quattro Concept technical information...... 18

3.1 Audi A6 Hybrid Technical Data[34]...... 26 3.2 Measured parameters, units and signal names...... 26 3.3 Approximations of the equivalent masses for each gear of the 8-speed gearbox...... 30 3.4 Simulink model extended gear ratios...... 31 3.5 Intervals for the energy vs power histograms...... 33 3.6 Values for c (α)...... 35 3.7 Example of offline priority table...... 40

4.1 Index i0 and i1 to be used for calculation of the electrical and fuel consumptions, for EV and HEV modes...... 55 4.2 Comparison between the three implementations of the global optimizer...... 57 4.3 Priority points for each drive mode and driving categories 2, 3 and 4...... 60 4.4 Summary of the three best drive strategies for the route 2-3-4-3-2-3-4-3-2...... 60 4.5 Summary of the four drive strategies for the route 2-3-4-3-2-3-4-3-2...... 62 4.6 Basic and enhanced basic drive strategies for the route 2-3-4-3-2-3-4-3-2...... 63 4.7 Fuel consumptions and savings between the different drive strategies for the route 2-3-4-3-2-3-4-3-2. 63

73 Nomenclature

α Slope angle

A Frontal area of the Audi A6

ACC Adaptive Cruise Control

ACC Adaptive Cruise Control

(αi)i=1..14 Conversion parameters for electrical consumption calculation

(βi)i=1..14 Conversion parameters for fuel consumption calculation BMS Battery Monitoring System

CDF Cumulative Distribution Functions

Cx Aerodynamic drag coefficient of the Audi A6

DOD Depth Of Discharge

∆E Energy increment

ECU Engine Control unit e Electrical consumption

EMS Energy Management System

EREV Extended Range Electric Vehicle

EUDC Extra Urban Driving Cycle

EV Electric Vehicle

~ex Unit vector along direction of motion

~ez Unit vector along local vertical direction, perpendicular to ~ex

F~L Aerodynamic resistance

F~r Total rolling resistance for both axles fr Rolling resistance coefficient FTP-75 EPA Federal Test Procedure f Fuel consumption

F~x Net tractive force

74 LIST OF TABLES LIST OF TABLES

~g Standard gravity GHG Greenhouse Gas GPS Global Positioning System HEV Hybrid Electric Vehicle HV High-Voltage ICE Internal Combustion Engine

Ie Engine, flywheel and clutch polar inertia

Ip Propeller shaft polar inertia

Iw Wheels and axles polar inertia

Je Engine, flywheel and clutch moment of inertia

Jp Propeller shaft moment of inertia

Jw Wheels and axles moment of inertia LEZ Low Emission Zone m Mass empty of the modified Audi A6 Hybrid mj Equivalent mass of the rotating parts

N~ Total ground reaction nengine Engine speed NEDC New European Driving Cycle OEM Original Equipment Manufacturer PDS Predictive Drive Strategy

PEM,max Electrical machine maximum output power at wheels PHEV Plug-In Hybrid Electric Vehicle

PICE,max ICE maximum output power at wheels

PICE+EM,max ICE and electrical machine (E-boost) maximum output power at wheels

Preg Absolute value of the maximum power at wheels that can be fully recuperated using regenerative braking

Px Net tractive power RDS Radio Data System rwheel Wheel radius of the Audi A6 SOC State Of Charge TMC Traffic Message Channel

Ufinal Final drive reduction ratio

st U1 1 gear ratio

75 LIST OF TABLES LIST OF TABLES

nd U2 2 gear ratio

rd U3 3 gear ratio

th U4 4 gear ratio

th U5 5 gear ratio

th U6 6 gear ratio

th U7 7 gear ratio

th U8 8 gear ratio UDC Urban Driving Cycle W~ Car weight

WLTP World-Harmonized Light-Duty Vehicles Test Procedure WPT Wireless Power Charging ZEZ Zero Emission Zone

76 Chapter 7

Appendix

7.1 Complementary Information about Audi 7.1.1 The car fleet

77 CHAPTER 7. APPENDIX 7.1. COMPLEMENTARY INFORMATION ABOUT AUDI

Figure 7.1: Overview of the Audi fleet 2013.

78 CHAPTER 7. APPENDIX 7.2. ALGORITHM IMPLEMENTATIONS

7.2 Algorithm implementations 7.2.1 Global optimizer: C-code 7.2.1.1 Tree construction: TreeBuilding.h and TreeBuilding.c

1 typedef struct Knoten Knoten; 2 struct Knoten 3 { 4 double pwr ; // Batterie Ladezustand [kWh] 5 double score ; // wie gut die Priorisierung respektiert wurde 6 double verbrauch ; // Kraftstoffverbrauch [kWh] 7 int *strategie; // Fahrmodi 8 Knoten *soehne; // Sohn 1 = HEV, Sohn 2 = EV, Sohn 3 = Charge 9 int index ; // Position in der Strecke 10 }; 11 12 Knoten neuerKnoten( double pwr , double score , double verbrauch , int *strategie, int index ); 13 int * naechsterKnoten(Knoten aktuellerKnoten, double * hev_elm , double * hev_vkm , double * ev_elm , double * ev_vkm , double *charge_elm, double * charge_vkm, int * segmenttyp, double * prio_hev , double * prio_ev , double * prio_charge, int n); 14 int updateBest(Knoten aktuellerKnoten, int n); 15 void copyvekt ( int * source , int * target , int n);

1 # include 2 # include 3 # include "TreeBuilding.h" 4 5 6 // Initialisierung 7 double KapaMIN = 0.2*10.8; 8 double KapaMAX = 0.8*10.8; 9 10 int * bestSTRATEGIE = NULL; 11 double bestVERBRAUCH = 1111111111; 12 13 Knoten neuerKnoten( double pwr , double score , double verbrauch , int *strategie, int index ){ 14 Knoten einNeuerKnoten; 15 einNeuerKnoten.pwr = pwr; 16 einNeuerKnoten.score = score; 17 einNeuerKnoten.verbrauch = verbrauch; 18 einNeuerKnoten.strategie = strategie; 19 einNeuerKnoten.index = index; 20 21 return einNeuerKnoten; 22 } 23 24 int * naechsterKnoten(Knoten aktuellerKnoten, double * hev_elm , double * hev_vkm , double * ev_elm , double * ev_vkm , double * charge_elm, double * charge_vkm, int * segmenttyp, double * prio_hev , double * prio_ev , double * prio_charge, int n){ 25 int index = aktuellerKnoten.index; 26 int newBest ; 27 int * bestStrat = NULL; 28 int * bestStrat0; 29 int * bestStrat1; 30 int * bestStrat2; 31 32 if(index>=n){ 33 newBest = updateBest(aktuellerKnoten,n); 34 if( newBest ){ 35 bestStrat = malloc(n* sizeof ( int )); 36 copyvekt(aktuellerKnoten.strategie,bestStrat,n); 37 } 38 return bestStrat ; 39 } 40 else { 41 double pwr = aktuellerKnoten.pwr; 42 double verbrauch = aktuellerKnoten.verbrauch; 43 double score = aktuellerKnoten.score; 44 int *strategie = aktuellerKnoten.strategie; 45 46 Knoten Knoten0; 47 Knoten Knoten1; 48 Knoten Knoten2; 49

79 CHAPTER 7. APPENDIX 7.2. ALGORITHM IMPLEMENTATIONS

50 // Sohn 0 : HEV 51 double pwr_next = pwr + hev_elm[index]; 52 if(prio_ev[segmenttyp[index]] == 5 || prio_charge[segmenttyp[index]] == 5){ 53 bestStrat0 = NULL; 54 } 55 else if(pwr_next>=KapaMAX){ 56 strategie[index] = 0; 57 Knoten0 = neuerKnoten(KapaMAX,score+prio_hev[segmenttyp[index]],verbrauch+hev_vkm[index], strategie,index+1); 58 bestStrat0 = naechsterKnoten(Knoten0,hev_elm,hev_vkm,ev_elm,ev_vkm,charge_elm,charge_vkm, segmenttyp,prio_hev,prio_ev,prio_charge,n); 59 } 60 else if(pwr_next=KapaMAX){ 75 strategie[index] = 1; 76 Knoten1 = neuerKnoten(KapaMAX,score+prio_ev[segmenttyp[index]],verbrauch+ev_vkm[index],strategie , index +1) ; 77 bestStrat1 = naechsterKnoten(Knoten1,hev_elm,hev_vkm,ev_elm,ev_vkm,charge_elm,charge_vkm, segmenttyp,prio_hev,prio_ev,prio_charge,n); 78 } 79 else if(pwr_next=KapaMAX){ 94 strategie[index] = 2; 95 Knoten2 = neuerKnoten(KapaMAX,score+prio_charge[segmenttyp[index]],verbrauch+charge_vkm[index], strategie,index+1); 96 bestStrat2 = naechsterKnoten(Knoten2,hev_elm,hev_vkm,ev_elm,ev_vkm,charge_elm,charge_vkm, segmenttyp,prio_hev,prio_ev,prio_charge,n); 97 } 98 else if(pwr_next

80 CHAPTER 7. APPENDIX 7.2. ALGORITHM IMPLEMENTATIONS

115 return bestStrat0; 116 } 117 } 118 return NULL; 119 } 120 121 int updateBest(Knoten aktuellerKnoten, int n){ 122 int k; 123 int newBest = 0; 124 double verbrauch = aktuellerKnoten.verbrauch; 125 if(verbrauch

1 int * getStrategie( double * HEV_ELM_Verbrauch, double * HEV_VKM_Verbrauch, double * EV_ELM_Verbrauch, double * EV_VKM_Verbrauch, double * CHARGE_ELM_Verbrauch, double * CHARGE_VKM_Verbrauch, int * Segmenttyp, double * Priorisierung_HEV, double * Priorisierung_EV, double * Priorisierung_Charge, int n); 2 double getPrioScore( int * strategie, int * Segmenttyp, double * Priorisierung_HEV, double * Priorisierung_EV, double * Priorisierung_Charge, int n); 3 double getKSV ( int * strategie, double * HEV_VKM_Verbrauch, double * EV_VKM_Verbrauch, double * CHARGE_VKM_Verbrauch , int n); 4 double * getELMverlauf( int * strategie, double * HEV_ELM_Verbrauch, double * EV_ELM_Verbrauch, double * CHARGE_ELM_Verbrauch, int n);

1 # include 2 # include 3 # include "TreeBuilding.h" 4 5 6 int * getStrategie( double * HEV_ELM_Verbrauch, double * HEV_VKM_Verbrauch, double * EV_ELM_Verbrauch, double * EV_VKM_Verbrauch, double * CHARGE_ELM_Verbrauch, double * CHARGE_VKM_Verbrauch, int * Segmenttyp, double * Priorisierung_HEV, double * Priorisierung_EV, double * Priorisierung_Charge, int n){ 7 8 double KapaINIT = 0.8*10.8; 9 double KapaMIN = 0.2*10.8; 10 double KapaMAX = 0.8*10.8; 11 12 int * bestStrat; 13 14 int * strat ; 15 strat = malloc(n* sizeof ( int )); 16 strat [n] = -1; 17 Knoten ersterKnoten = neuerKnoten(KapaINIT,0.,0.,strat,0); 18 bestStrat = naechsterKnoten(ersterKnoten,HEV_ELM_Verbrauch,HEV_VKM_Verbrauch,EV_ELM_Verbrauch, EV_VKM_Verbrauch,CHARGE_ELM_Verbrauch,CHARGE_VKM_Verbrauch,Segmenttyp,Priorisierung_HEV, Priorisierung_EV,Priorisierung_Charge,n); 19 20 return bestStrat ; 21 } 22 23 double getPrioScore( int * strategie, int * Segmenttyp, double * Priorisierung_HEV, double * Priorisierung_EV, double * Priorisierung_Charge, int n){ 24 double score = 0.0; 25 double delta_score; 26 int k; 27 for (k=0;k

81 CHAPTER 7. APPENDIX 7.2. ALGORITHM IMPLEMENTATIONS

31 else if(strategie[k]==1){ 32 delta_score = Priorisierung_EV[Segmenttyp[k]]; 33 } 34 else if(strategie[k]==2){ 35 delta_score = Priorisierung_Charge[Segmenttyp[k]]; 36 } 37 score = score+delta_score; 38 } 39 return score ; 40 } 41 42 double getKSV ( int * strategie, double * HEV_VKM_Verbrauch, double * EV_VKM_Verbrauch, double * CHARGE_VKM_Verbrauch , int n){ 43 double KSV = 0.0; 44 double delta_KSV ; 45 int k; 46 for (k=0;kKapaMAX){ 73 ELMverlauf[k] = KapaMAX; 74 } 75 } 76 else if(strategie[k-1]==1){ 77 ELMverlauf[k] = ELMverlauf[k-1]+EV_ELM_Verbrauch[k-1]; 78 if(ELMverlauf[k]>KapaMAX){ 79 ELMverlauf[k] = KapaMAX; 80 } 81 } 82 else if(strategie[k-1]==2){ 83 ELMverlauf[k] = ELMverlauf[k-1]+CHARGE_ELM_Verbrauch[k-1]; 84 if(ELMverlauf[k]>KapaMAX){ 85 ELMverlauf[k] = KapaMAX; 86 } 87 } 88 } 89 return ELMverlauf; 90 } 7.2.1.3 Main: main.c

1 # include 2 # include 3 # include "TreeAnalysis.h" 4 5 int main ( int argc , char * argv []) { 6 7 //Input values 8 double CHARGE_VKM_Verbrauch[3] = {123.733014440796,154.625076323098,158.272660544230}; 9 double EV_VKM_Verbrauch[3] = {0.,0.540135260032777,70.0907910731235}; 10 double HEV_VKM_Verbrauch[3] = {4.48083944035603,14.3735520891996,78.3656083421454}; 11 double CHARGE_ELM_Verbrauch[3] = {1.07529399207061,1.53189589648340,1.38246675579402}; 12 double EV_ELM_Verbrauch[3] = {-2.42738374147590,-5.65939179522888,-1.54572289904641};

82 CHAPTER 7. APPENDIX 7.2. ALGORITHM IMPLEMENTATIONS

13 double HEV_ELM_Verbrauch[3] = {-0.641506792786493,0.293538284077632,1.89047047679850}; 14 15 int Segmenttyp[3] = {1,2,3}; 16 17 double Priorisierung_HEV[8] = {1.,0.,1.,2.,0.,0.,0.,0.}; 18 double Priorisierung_EV[8] = {2.,2.,1.,0.,2.,0.,0.,0.}; 19 double Priorisierung_Charge[8] = {0.,1.,2.,1.,0.,0.,0.,0.}; 20 21 int n = sizeof (Segmenttyp)/ sizeof ( int ); 22 int * bestStrat; 23 24 bestStrat = getStrategie(HEV_ELM_Verbrauch,HEV_VKM_Verbrauch,EV_ELM_Verbrauch,EV_VKM_Verbrauch, CHARGE_ELM_Verbrauch,CHARGE_VKM_Verbrauch,Segmenttyp,Priorisierung_HEV,Priorisierung_EV, Priorisierung_Charge,n); 25 26 printf("Segmenttyp :\t"); 27 int k; 28 for (k=0;k

83 CHAPTER 7. APPENDIX 7.2. ALGORITHM IMPLEMENTATIONS

7.2.2 Global optimizer: ASCET

.

Figure 7.2: Overview of the predictive drive strategy algorithm in ASCET

84 CHAPTER 7. APPENDIX 7.2. ALGORITHM IMPLEMENTATIONS

Figure 7.3: Predictive drive strategy optimizer in ASCET. 85 CHAPTER 7. APPENDIX 7.2. ALGORITHM IMPLEMENTATIONS

Figure 7.4: Predictive drive strategy consumption unit in ASCET.

86