Portal de Periódicos da CAPES

Optimisation of energy efficiency based on average driving behaviour and driver's preferences for automated driving

2014; Institution of Engineering and Technology; Volume: 9; Issue: 1 Linguagem: Inglês

10.1049/iet-its.2013.0121

1751-9578

Philipp Themann, Julian Bock, Lutz Eckstein,

Human-Automation Interaction and Safety

IET Intelligent Transport SystemsVolume 9, Issue 1 p. 50-58 Regular PaperFree Access Optimisation of energy efficiency based on average driving behaviour and driver's preferences for automated driving Philipp Themann, Corresponding Author Philipp Themann [email protected] RWTH Aachen University – Institut für Kraftfahrzeuge (ika), Steinbachstr 7, Aachen, 52074 GermanySearch for more papers by this authorJulian Bock, Julian Bock RWTH Aachen University – Institut für Kraftfahrzeuge (ika), Steinbachstr 7, Aachen, 52074 GermanySearch for more papers by this authorLutz Eckstein, Lutz Eckstein RWTH Aachen University – Institut für Kraftfahrzeuge (ika), Steinbachstr 7, Aachen, 52074 GermanySearch for more papers by this author Philipp Themann, Corresponding Author Philipp Themann [email protected] RWTH Aachen University – Institut für Kraftfahrzeuge (ika), Steinbachstr 7, Aachen, 52074 GermanySearch for more papers by this authorJulian Bock, Julian Bock RWTH Aachen University – Institut für Kraftfahrzeuge (ika), Steinbachstr 7, Aachen, 52074 GermanySearch for more papers by this authorLutz Eckstein, Lutz Eckstein RWTH Aachen University – Institut für Kraftfahrzeuge (ika), Steinbachstr 7, Aachen, 52074 GermanySearch for more papers by this author First published: 01 February 2015 https://doi.org/10.1049/iet-its.2013.0121Citations: 18AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Share a linkShare onFacebookTwitterLinkedInRedditWechat Abstract The implementation of anticipating driving styles in adaptive cruise control systems promises to considerably reduce fuel consumption of vehicles. As drivers have to accept the optimised driving styles of such systems, which implement longitudinally automated driving, the optimisation results should not deviate strongly from the average driving behaviour. This work presents an approach to the optimisation of the vehicle's longitudinal dynamics, which is based on a predicted average driving profile. The proposed approach ensures that the optimisation results meet the expectations of drivers by directly accounting for driver's preferences on weighting up travel time against fuel consumption relative to the average driving profile. Based on human decision finding, rational and intuitive planning decisions are modelled in a cost function and represent optimisation constraints. The approach generally includes information from vehicle-to-vehicle and vehicle-to-infrastructure communication (V2X), which is an extension to the state-of-the-art. This study describes the optimisation approach and presents a method to determine suitable optimisation parameters in order to consider driver's preferences. The optimisation approach is applied in a simulated test drive and improvements in fuel economy are analysed. Finally, the authors sketch a reference system architecture to prove the feasibility of the presented approach. 1 Introduction and motivation Currently, the European Commission (EC) challenges automobile industry by setting a limit of 130 g CO2 per km for a fleet average to be achieved in 2015, which is planned to be lowered to a limit of 95 g CO2 per km in 2020 [1]. The automobile industry has to face these challenges to find solutions to lower CO2 emissions significantly. As CO2 emission is directly proportional with fuel consumption, energy efficiency of vehicles gains crucial importance. Predictive driving provides great potentials in lowering fuel consumption [2]. Professional driver trainings coach drivers to implement predictive driving styles, while the potential savings in fuel consumption diminishes over time [3]. Consequently, driver assistance systems are useful to support the driver incorporating fuel-saving driving styles without decreasing performance. To completely exploit the fuel-saving potential of predictive driving, a driver assistance system needs to control the longitudinal dynamics of a vehicle. Thus information not directly available for drivers is evaluated and energy-efficient velocity trajectories are realised. Besides information from vehicle sensors, digital map data, vehicle-to-vehicle and vehicle-to-infrastructure communication (V2X) are also used within the approach presented here. Based upon this information, the assistance system calculates a speed profile, which is optimal with respect to fuel consumption. This speed profile possibly differs significantly from the uninfluenced driving style and the driver might thus refuse using the system. If this results in switching off the system, improvements in fuel economy diminish. Thus, a driver assistance system with automatic longitudinal control needs to consider driver acceptance in the optimisation of a speed profile, which is elaborated in the following sections of this work. 2 State-of-the-art of optimisation algorithms in adaptive cruise control (ACC) systems Systems for the control of the vehicle's longitudinal velocity are available in the market for quite some time. Already in 1958, Chrysler introduced a system in the Imperial, which automatically controls the velocity to the value requested by the driver [4]. This 'Auto-Pilot' is the basis for many cruise control systems available today. Cruise control has originally been designed as a system aiming to purely increase comfort. As side effects, fuel efficiency and safety of vehicles are improved as the vehicles are driven in constant operation modes [4]. As a consequent improvement of cruise control, the so-called adaptive cruise control (ACC) systems have been developed. The ACC systems support drivers especially on motorway and interurban drives by controlling both the distance to front vehicles and also the velocity. The further development of cruise control was mainly driven by the increase in performance of sensor technologies. Current sensors enable the ACC systems to reliably determine the distance and the relative velocity to front vehicles. In future, additional information from V2X technologies will be available and hence enable the further development of current ACC systems, especially with respect to fuel economy. This information is a precondition to automate the longitudinal dynamics of the vehicle. The ACC systems implementing an optimisation in gear-shift behaviour of heavy goods vehicles are already on the marked. Map data is used to consider topography in the calculation of optimal gear shifts in ACC systems. Besides research activities [5], original equipment manufacturers and suppliers such as Scania [6], Daimler [7] or ZF Friedrichshafen [8] also provide similar systems that can improve fuel efficiency by 2–4%. However, these systems are not able to consider dynamic information from V2X technologies and mainly rely on static map data. Some research projects elaborate on energy-efficient ACC systems. Volkswagen presented an energy-efficient ACC system in a research project in 2012 [9]. It uses data from a digital map to replace braking by energy-efficient deceleration strategies like coasting in neutral or fuel cut-off. Driver input is considered in the determination of alternative deceleration strategies and the system realises improvements in fuel efficiency by 13% [9]. Similarly, at Porsche the so-called 'InnoDrive' system uses map data to optimise the velocity profile. This system realises improvements in fuel efficiency by 10% [10] and is based on an optimisation of the speed profile, which is presented in [11]. The system uses the method of dynamic programming for optimisations. An optimal speed profile is calculated in real-time while driving, but is purely based on static map data. Different driving styles, which can be chosen by the driver, are achieved by change of parameters in a cost function. The optimisation of a speed profile, which uses a cost function with linear weighting between trip length and fuel consumption, is presented in [12]. Dynamic programming is also used for optimisations to deploy fuel savings based on the knowledge of future traffic light status information. However, the optimisation is not applied in a vehicle yet. For a real-time capable implementation of dynamic programming, calculation power available in the vehicle is crucial. To overcome the limitations resulting from insufficient calculation power in vehicles, some research proposes to pre-calculate simple action rules based on offline optimisations with dynamic programming [13]. The research approaches mentioned above already propose some optimisation algorithms but still are limited with respect to the consideration of information from V2X technologies. The approach presented in [14] predicts the behaviour of the first vehicle in front and assumes that the influence of additional front vehicles is approximated by the status of the immediate front vehicle. A simple fuel-consumption model is used to implement a model predictive control. The system is validated in simulations for approaches of a traffic light, while the cost function used in the optimisation is not analysed in detail. Reductions in fuel consumption by 13% are reported, while travel time increases by 16% [15]. Generally, the approach hence is promising with respect to the incorporation of information from V2X communication, although it lacks a method to define suitable weighting factors for the cost function in the optimisation. Additionally, the restricted prediction horizon considering just the immediate front vehicle and the simplifications in the modelling of fuel consumption are critical. Several approaches presented in the literature apply dynamic programming to optimise energy efficiency of vehicles without considering travel time or driver's preferences directly. In [16], the cost function, which is optimised, purely considers the vehicle's energy consumption. The optimisation approach chooses a velocity of approximately 70 km/h on a motorway section, whereas the reference profile is at more than 120 km/h. The resulting high potential to improve energy efficiency in this situation will hardly be accepted by drivers. In [17], a more complex cost function is used in the optimisation that considers the state of charge of the battery in a hybrid vehicle and additionally the travel time. A suitable weighting factor to properly consider the battery's state of charge in the dynamic programming is identified but driver's preferences are not explicitly modelled. 3 Research methodology and optimisation approach The research approach described in this work is deduced from the state-of-the-art analysis summarised in the previous section. The research projects currently do not fully exploit the predictive driving approach. Either information available by V2X communication is not used or alternative energy-efficient speed profiles are only calculated for specific driving situations and not for a longer distance in front of the vehicle. Further gaps are lacking on implementations of optimisation approaches in vehicles and the insufficient adaptability of system behaviour to different driver preferences. Within the scope of this work, the term 'driver preferences' is used to express the driver's tendency to minimise either travel time or fuel consumption relative to the average driving profile. This work approaches the gaps described above, and consequently, requirements on the research approach are derived. An energy-efficient ACC system presented here automates the vehicle's longitudinal dynamics and has to cope with the following challenges along with the corresponding requirements: Expand conventional ACC functionalities by optimisation of energy efficiency: The system needs to deploy energy-efficient driving strategies such as 'coasting' or 'fuel cut-off' either automatically or by recommending these to the driver. Consideration of driver's preferences: The driver should be able to adjust the system behaviour with just one input value to tune the system from maximum energy efficient to minimal deviation from average driving behaviour. Exploitation of digital map data incorporating information from V2X communication and implementation on a test vehicle without restrictions on test routes: The system architecture needs to ensure connectivity to different data sources and a real-time capable implementation. This paper especially focuses on the consideration of driver's preferences in the optimisation. The characteristics of the optimisation approach, which fulfils the requirements listed above, are defined by two modules: 1. Prediction module: The average driving behaviour is predicted for the upcoming traffic situations. • Inputs: Data from a digital map (e.g. distance to slopes, curves, speed limits, traffic lights), from sensors (e.g. distance to front vehicles and their velocity) and also from V2X communication (e.g. time to green of traffic lights, position of other vehicles in the vicinity). • Output: A vector representing predicted velocities for average driving behaviour against the distance ahead of the vehicle (for each distance step of e.g. 10 m in front of the vehicle, a corresponding velocity is calculated). The vector of predicted velocities is provided both for the vehicle hosting the system and also for front vehicles that are detected, for example, by radar sensors or via V2X communication. 2. Optimisation module: The optimisation algorithm assesses the utility of driving strategies relative to the average driving behaviour. A utility function contains a weighting factor that expresses the driver's preferences on weighting up travel time against fuel consumption. • Input: Data from a digital map, the predicted velocity profile and the characteristics of the vehicle represented in a look-up table (see sections below). • Output: A vector representing the velocities, which are optimised with respect to the drivers preferences, against the distance ahead of the vehicle. These velocities are either used to automatically control the vehicle or to recommend certain driving manoeuvres to the driver. In the following, first of all the prediction module is described in a separate section as this is a necessary precondition for the optimisation approach, which the following sections elaborate on. 4 Prediction module: anticipation of driving behaviour as a reference To assess the velocity of the vehicle for an upcoming situation ahead, a prediction model is used. For the scope of this work, the so-called 'ecoSituational Model' (eSiM) prediction model is used, which is developed in the research project eCoMove [18]. The eSiM model provides a short-term prediction in the form of a velocity profile against distance. This velocity profile is used as a basic input for the optimisation to derive suitable driving strategies minimising fuel consumption on the road ahead. Furthermore, the prediction model provides a classification between current and predicted driving situations, which can be used to consider situation-specific preferences of drivers with respect to different driving strategies. Different information sources such as radar sensors, vehicle-to-vehicle or vehicle-to-infrastructure communication are evaluated by the model, which enables a cooperative prediction [19]. To predict the velocity of a vehicle, the different entities of traffic need to be considered in the prediction model: environment, driver and vehicle. Each entity has an impact on the velocity profile which the driver chooses in a specific traffic situation. The environment contains static and dynamic information about external influences on the driver and vehicle. Static information includes all information about the road such as slopes, curvatures or speed limits, whereas dynamic information represents traffic jams, construction sites or obstacles. In addition to the variation in environment, the driver of a vehicle can vary in driving behaviour or driving mood. A sporty driver, for example, results in totally different velocity profiles than a conservative driver. The third entity affecting the chosen velocity profile is the vehicle itself. Technical aspects such as total vehicle mass, drive train performance or aerodynamic resistances heavily affect the acceleration of the vehicle. A passenger car has different dynamics than a heavy commercial vehicle. The prediction results from the model have been verified in test drives, which prove the feasibility [20]. 5 Optimisation module: derive driving strategy based on reference behaviour According to the requirements defined based on the analysis of the state-of-the-art, the optimisation module identifies the driving strategy for the upcoming traffic situation, which is best with respect to the driver's preferences. The driving strategies in this context are characterised by a velocity profile along with the upcoming road section and the drive train operation mode. Thus, the driving strategies such as coasting, roll-out or fuel cut-off are also need to be considered by the optimiser. These strategies are defined by the corresponding drive train operation mode, and the velocity profile implicitly results from this. The following section first of all describes the utility function the optimiser considers when assessing the usefulness of different driving strategies for a road section ahead. 5.1 Utility function modelling human preferences An approach to model human decision processes with respect to the choice of energy-efficient driving strategies is presented in [21] and implemented in the optimisation approach discussed in this work. Research in cognitive and social psychology suggests that two different kinds of processes have to be considered in human decision making, which is called dual-process theory. On one hand, there are 'system 1' processes that are classified as relatively fast, automatic and unconscious. On the other hand, there are 'system 2' processes that are slow, effortful and conscious [22]. In the context of energy-efficient driving, 'system 1' processes in human decision making reflect situation-dependent expectations on the driving behaviour of the vehicle. In a deceleration situation, in front of a traffic light for example, a roll out strategy would reduce fuel consumption of the vehicle, but on the other hand could result in a slow approaching, which might be rejected by the unconscious and intuitive decision in system 1 processes of the driver. To consider these decision aspects in the optimisation, the deviation of the optimised velocity from the predicted average velocity is limited. This way the system defines acceptable velocities along the distance ahead, which are called driving corridor in the following. The driving corridor is based on the prediction results and used as a restriction in the definition of the state space used for optimisation. Consequently, optimised velocities of the system are all within the driving corridor and do not deviate heavily from the average driving. The maximum and minimum accepted velocities define the width of the driving corridor, which determines the size of the state space of all allowed combinations of velocity and distance. System 2 processes in human decision making reflect the outcome of slow, effortful and conscious decisions, weighting up the advantages and the disadvantages of pursuing an energy-efficient strategy in a traffic situation. To reflect these processes, the system considers the two attributes energy efficiency (or fuel consumption) and travel time in a linear additive utility function. The two utilities improvement in energy efficiency uE and reductions in travel time uT are combined with the weighting factor wE and define the total utility u2, which is defined in (1). (1) The parameter wE ∈ [0, 1] is used to weight relative fuel consumption against travel time. In the system, the weighting parameter wE is set by the driver to choose between close to average driving (wE = 0) and energy-efficient driving (wE = 1). This allows deriving utility values u2 for different driving strategies and choosing the strategy, which represents the driver's preferences best, by comparing these values. The utility uE ∈ [0, 1] assesses the improvement in energy efficiency of a driving strategy relative to the average driving behaviour. Hence, the relative reduction of fuel consumption frel of the optimised (eco) driving strategy feco relative to the average driving strategy favg is useful in determining the utility uE. Average driving in this context is characterised for each driving situation by the output of the prediction module described above. As the vehicle can deploy different driving strategies, these result in different relative reductions in fuel consumption . For the driving strategy with highest relative fuel consumption reduction , the utility value should be maximal: . In the same driving situation, the strategy with minimal fuel consumption reduction results in a minimal utility value: . Equation (2) realises an exponential transformation of the relative fuel consumption reduction frel to derive the utility value uE for a driving strategy. The parameter cE allows to further shape the course of the utility value over the relative fuel consumption reduction. (2) According to the requirements of the system mentioned above, the driver should be able to use just one input value in order to adapt the system behaviour. This intends to ensure the ease of use of the system. The impact of the parameter cE on the total utility function is not as important as the impact of the weighting factor wE. For cE → 0, the notation in (2) converges into a linear equation given in (3). (3) Similarly, to the utility uE also the utility uT is defined. The relative reduction in travel time trel compares the travel time tavg of the average driving behaviour to the travel time teco of optimised (eco) strategies. For different driving strategies, there are varying relative travel time reductions . This range of reductions is used in (4) to define the exponential utility function uT (trel). Owing to the same considerations mentioned above, the parameter cT is neglected and the utility function converges for cT → 0 is neglected and the utility funct (5). (4) (5) The output of the prediction module defines the travel time tavg and the fuel consumption favg of the predicted average driving behaviour. Hence, the utilisation of the prediction model is a precondition to use this approach. The unaffected average driving behaviour of the human driver is chosen as a reference. All energy-efficient driving strategies are compared relatively to this by the chosen utility function. Merging (3) and (5) into the definition of the utility function in (1) reveals the impact of the relative reductions in fuel consumption and travel time of the optimised strategy compared to the average driving. The result is given in (6) and transformed into (7) using the definitions from (2) and (4) (6) (7) The optimisation algorithm compares the utility value u2 of different driving strategies and chooses the strategy with highest utility value. The constants summed up or subtracted in the utility function can hence be neglected without affecting the result of the optimisation algorithm. Similarly, the constants that represent a factor of the total utility term do not affect the optimisation. The transformation of (7) into (8) reveals a constant term that is consequently not considered in the following. (8) Furthermore, the multiplication of the utility function with a constant factor does not affect the result of the optimisation algorithm. Based upon (8), thus a notation of the utility function to be used in the optimisation algorithm u2,opt is derived, which is given in (9). The total utility u2,opt results from a subtraction of a term from the constant kconst. (9) The term that is subtracted from the constant kconst represents costs for the optimised driving strategies compared to average driving. The costs of the strategy result from a linear combination and contains the constant factor kband as well as the weighting factor wE. For a weighting factor wE = 1, the cost term in (9) is represented by the quotient of the fuel consumptions and hence the driver's intention is considered appropriately. The constant factor kband represents a normalisation factor taking into account the different value ranges for travel time teco and fuel consumption feco of the optimised strategies. This factor kband thus describes the bandwidth of the strategies and ensures an equal distribution of the weighting parameter wE along with the different alternatives of driving strategies. The factor kband varies for different traffic situations and also depends on the implementation concept of the optimiser. In a below section, the factor kband is identified from simulations with the chosen optimisation concept. 5.2 Implementation concept of a suitable optimisation algorithm The optimisation algorithm needs to identify a driving strategy with minimal costs for the upcoming road section. Furthermore, it is required to be real-time capable, as it should be executed in a vehicle while driving. Real-time computing thereby has to be understood as correct computation finished in time. A discrete representation of the speed profile is used in order to fulfil these requirements. The optimisation of one single parameter per route segment results from the discretisation of driving routes, which is thus a multi-stage decision process. Furthermore, discrete values for the velocity of the vehicle are used. The vehicle's dynamics is represented by a set of discrete accelerations and decelerations that are called longitudinal dynamics variants in the following. The longitudinal dynamics variants also include energy-efficient deceleration strategies such as coasting in neutral and fuel cut-off as well as discrete values for accelerations and braking. The choice of the discrete acceleration and deceleration values determines the number of longitudinal dynamics variants that need to be considered in the optimisation. As an example, the discretisation steps for accelerations of 0.1 m/s2 are used within a range from −4 to +4 m/s2 (40 variants for acceleration and 40 variants for deceleration) and additionally the strategies coasting in neutral, coasting in gear, fuel cut-off and also constant driving sum up to 84 longitudinal dynamics variants. These discrete values define the state space for optimisation as a directed graph with the values of a cost function defining edge weights. The edges represent possible speed profiles of longitudinal dynamics variants, whereas the nodes are velocity states at discrete route points. Given a certain starting velocity at the starting position and a target velocity at the target position, the optimisation problem is now to find the path with lowest costs between the starting and the target node. In graph theory, this problem is called single-pair shortest path problem, which is often solved by Dijkstra's algorithm. Dijkstra's algorithm [23] is an application of the principle of dynamic programming on the shortest path problem [24]. Dynamic programming is based on Richard Bellman's principle of optimality: 'An optimal policy has the property that whatever the initial state and initial decision are, the remaining decisions must constitute an optimal policy with regard to the state resulting from the first decision.' [25]. In the context of shortest path problem this means that a shortest path between two nodes A and B, which follows across nodes M and N, also follows across the shortest path between M and N. The problem representation by a graph allows further techniques to reduce optimisation time. The results, which need to be calculated by a vehicle model and are used to evaluate the cost function, can be pre-calculated offline and saved in a look-up table. This is valid since only discrete states are considered in the graph. The complexity of evaluating the cost function can be kept low to a few memory accesses and simple arithmetic operations. By that, simulations with a vehicle model are not necessary while driving and computational demand is reduced significantly. Furthermore, calculation time, which directly depends on the number of edges and nodes, can be decreased by reducing the number of these edges and nodes. The graph size is determined by suitable boundary conditions, which represent a driving corridor and are explained in the subsections below. Using these techniques and Dijkstra's algorithm, the requirements mentioned above are met. The implemented approach guarantees to find a global optimum and the chosen methods allow real-time computing. Loss of precision by discrete values is a drawback, but for performing real-time optimisation a certain loss of precision by simplifications cannot be avoided. A higher precision is likely to be achieved in future by first of all increasing computing power. 5.3 Look-up table and discretisation parameters To make use of an offline pre-processing, fuel consumption, driving time and target velocity are calculated for a distance Δs depending on the starting velocity. The results of this calculation are stored in a look-up table. The calculation is performed with a vehicle model, which is based on the determination of driving resistances. The look-up table needs to fully cover the expected range of the most impo