WLTP Random Cycle Generator

Transcription

1 WLTP Random Cycle Generator Dennis G. Kooijman, Andreea E. Balau, Steven Wilkins Powertrains TNO Helmond, the Netherlands Norbert Ligterink, Rob Cuelenaere STL TNO Delft, the Netherlands Abstract European light duty vehicle emission legislation is gradually shifting the focus from test procedures with merely static test cycles, towards procedures including Real Driving Emissions (RDE), as they are a mean to achieve the European (NOx) emission reduction target. Hence a RDE trip must represent European driving behavior, such that real world driving emissions can be assessed. However, the usage of portable emission measurement systems has technical limitations. An alternative is testing on the chassis dynamometer in the laboratory. In order to cover a wide range of driving behavior to evaluate the real world driving emissions, a Random Cycle Generation methodology based on two Markov chains and real world measured data is developed. This methodology is used in order to develop a WLTP Random Cycle Generator Tool. Keywords drive cycle, random, WLTP, emissions, tool, vehicle cycle, driving behaviour, stochastic generator I. INTRODUCTION To assess the functionality and performance of vehicle (powertrain) systems, and parts thereof like emission control and OBD monitors, it appears to be necessary to test vehicles under not-predefined randomized conditions [1]. In order to evaluate the functioning of these systems under real world driving conditions, the EU RDE-LDV working group is developing a method based on the use of portable emission measurement systems (PEMS). However the use of PEMS equipment is problematic due to the technical limitations of the measurement of some pollutants under real driving conditions, such as particle numbers (PN) required by European Regulation 459/212/EC [2]. An alternative to PEMS measurements on vehicles may therefore be to perform tests on a chassis dynamometer under randomized conditions, like randomized driving cycle which must: reflect and cover European driving dynamics with respect to speed and acceleration profiles on the road, be drivable by the driver and the vehicle under test, meet requirements that are set by the current RDE legislation test under development, like road-type ratio s, cycle distance, cycle time. To this end, a methodology and software tool needs to be developed, in order to generate random cycles that can be driven by a light duty vehicle on a chassis dynamometer, and can be used for assessing criteria pollutant emission control for the type approval of light-duty vehicles. Other test conditions, such as ambient temperature and humidity of the driving cycle may also be randomized, but are here out of scope. In literature multiple approaches for generating driving cycles can be found: manual short trip (i.e. driving events between two idling periods) concatenation of recorded data in a trip database, such that specific chosen cycle characteristics meet the desired characteristics that represent the average driving [3], or short trip concatenation [5] using evolutionary algorithms [4]. Other approaches use traffic models [6], where others follow more stochastic approaches, using Markov chains, bases on trip distance [7], [8], [9] or road types [1]. This work proposes a two-stage Markov chain approach for generating new drive cycles, where first a sub road-type (on measurement optimized velocity band within a road-type) is stochastically generated on top of the legislation prescribed road type profile. The second stage is the generation of second by second accelerations, based on the probabilities also derived from measurement data. This creates a more random and wider coverage of speed-acceleration states compared to the short-trip approaches, and adds flexibility to adjust the cycle for vehicle limitations during generation. Using road-types covers the assumption that the driver will not behave the same when transiting a rural road or when driving on a motorway; as such, the acceleration will have a statistically different behavior, and vehicle (engine) dynamics will be different for different speeds. A further advantage of using a sub-road type approach is that it increases the steady driving situations, while maintaining the variability within a speed-band, all with the same probability distribution as taken from measurement data. In other applications, this method can be used for load cycle prediction if augmented with routes from navigation equipment or route prediction [11]. For the development of this methodology the European set of the WLTP database was used as input data for both Markov route generation, as the acceleration Markov chain. This database was created with the help of a number of European countries that collaborated and provided real world driving measurements for development of the WLTC. Consequently

2 cycles that contain European typical accelerations per velocity and road types are generated, such that these cycles are representative to real driving behavior. The paper content is structured as follows: first, the Methodology for the development of the random cycle generation will be presented, remarking some of the design decisions and how the general problem was approached. Second, the WLTP Random Cycle Generator Tool will be presented including the options available to the user and the generated results. Next, the Results and Validation section will present and validate the tool generated driving cycles. Finally a Conclusions section is also included. II. METHODOLOGY The stochastic drive cycle generation methodology is developed in Matlab and is based on Markov processes. Markov theory for stochastic behavior has been extensively studied, and the main references regarding the theoretical background on the Markov chains used in this paper are [12] and [13]. Two separate stochastic generators are used: one for generating the road type and one for generating the vehicle acceleration. In the first stage of the approach, a stochastic generator is used to create a semi-random route profile for the three different road types that are considered in the WLTP database: urban, rural and motorway. Each road type is further divided into sub-road types corresponding to different velocity bins. The WLTP RCG methodology follows the principles from on-road RDE testing. Hence, the restriction in driving behavior to the WLTP database is to be combined with a representative distribution of urban, rural, and motorway driving. This requires a large shift in WLTP RCG methodology from timebased [1] to distance-based cycle generation. The original distance distribution was 34% - 33% - 33%, however, a ±5% randomly allowed tolerance is implemented. Moreover, as the definition of urban, rural and motorway is fuzzy, certainly across Europe, the split is based on velocity characteristics. Also this method for categorizing the road type is still partly open. For example excluding stopping times from the averaging, or using maximal velocity (between stops) instead of average velocity will result in minor adaptions of the roadtype designation, however, with possibly further changes in the WLTP RCG methodology. In the second stage, a stochastic generator is used to generate the vehicle acceleration corresponding to the already generated road type profile. A stop is imposed at the end of the cycle. The methodology used in order to generate random driving cycles representative for real driving behavior is more extensively described in the following steps: A. Road type distance percentage generation The total cycle distance and the order of the road types is fixed: urban-rural-motorway, with a base percentage of 34% - 33% - 33% respectively, and a ±5% randomly allowed tolerance. In order to generate the random distance percentages of each road-type, three steps need to be taken: 1. A random percentage deviation is generated for the urban road-type distance: a random number (R Urban ), between -5 and 5 is generated, representing the deviation for the 34% urban road-type distance. 2. Considering the generated percentage deviation of the urban road-type, a random percentage deviation is generated for the rural road-type distance: a random number (R Rural ), between max(- 5, (-5-R Urban )), and min(5, 5- R Urban )) is generated, representing the deviation for the 33% rural roadtype distance. 3. Third, the percentage of motorway road-type distance is calculated, as the remaining distance. B. Route generation The first stochastic generator is used to generate a semirandom route profile corresponding to the three different road types: urban, rural and motorway. In order to generate the semi-random road type profile, the following steps need to be taken: 1. Data processing: First, the data used in the stochastic generators has to be collected and processed. For the development of the World-wide harmonized Light duty Test Cycle (WLTC), driving data was collected from several countries. Velocity profiles for a number of vehicles and a number of days of driving are recorded in this database. In some cases velocities and accelerations were not physically feasible, e.g. velocities above 3 km/h or accelerations above 5 m/s2. The data was therefore smoothed with a Hamming filter and grouped into sub-cycles from stop to stop according to the maximal velocity. This data has been made available by the European Commission for generating the input probabilities for the random cycle generator. The data was not further manipulated, except for filling in gaps in between the different sub-cycles. As the central idea behind the WLTP Random Cycle Generator is the second-by-second changes in accelerations for a given velocity. If a person accelerates, there is a high probability to remain accelerating. This also depends on the magnitude of acceleration. These probabilities are determined from the WLTP database, and they form the basis of the second-by-second changes in the generated cycle. The average velocities over periods of about a minute are also determined. They are called "subroad-types" but are actually a characterization of the average velocities. They are grouped into average velocity bands, likewise those of the WLTP database (which is originally based on peak velocities between two stops), and these are

3 the three road-types (urban, rural and motorway) visible in the generated cycle. Fig. 1. Road types and sub-road types 2. Probability matrix generation: The probability of a transition from one sub-road-type to the next is also determined from the WLTP database, and they are the probabilities of the minute-by-minute transitions in the generator. With these two sets of probabilities we span the short-time dynamics and the medium-time dynamics. For each road type (urban, rural, motorway) probability matrices for transitions from each subroad-type to sub-road-type are being generated (P SRT SRT Sub-Road Type to Sub-Road Type Probability matrix). 3. Route generation: The total cycle distance is known, and the percentage of each road-type was generated at the pervious point of the algorithm. Now, for each road type, random sub-road types will be generated, on a second-by-second base, using the P SRT SRT probability matrix. At this point, we have the generated route profile. C. Speed generation The second random generator function is used to generate the vehicle acceleration corresponding to the already generated road type profile. 1. Data processing: First, for each sub-road-type, speed data and acceleration data will be used to partition the data into classes and states. The states are related to acceleration and speed profiles, while the classes are related solely to the velocity profile. Fig. 2. Speed classes and acceleration states 2. Probability matrix generation: For each speed class, a probability matrix is then generated including transitions between acceleration states (P Acc Acc Acceleration to Acceleration Probability matrix). 3. Acceleration generation: On a second-by-second base, for the current sub-road-type, a corresponding acceleration is generated based on the P Acc Acc probability matrices. 4. Vehicle limitation check: The amount of high acceleration is limited, which seems to suggest that all countries and participating institutes have selected average, medium-powered vehicles. Likewise, the generated driving behavior may be on the same side for a high-powered vehicle. However, the driving behavior may require occasionally too much power to be achievable by a low power or high mass vehicle. This yields an important restriction on the generated cycle based on vehicle characteristics. In order to circumvent this obstacle, the drive cycle generator cross-check the speed profile against the vehicle characteristics: the vehicle total mass, an average rolling resistance force, velocity-dependent resistance force, air drag force, maximum allowed deceleration and the rated engine power. This will limit the maximum allowed acceleration of the vehicle, making it representative to a specific vehicle type. 5. Speed generation: From the generated acceleration, the corresponding speed profile is derived. At this point, we have the generated speed profile. D. Stop initiation At the end of the cycle, a forced stop is imposed, by setting the sub-road type road type to the one representative for the lowest speed bin. The corresponding acceleration will be randomly generated, allowing a natural stop behavior. E. Gear shift advice generation Based on the generated speed profile, a gear shift advice is calculate, using four different methods: HS, CADC, NEDC and DNEDC. An advice on when to change gears is generated, in order to follow the generated speed profile. Some vehicle parameters are required: the vehicle total mass, an average rolling resistance force, velocity-dependent resistance force, air drag force, maximum allowed deceleration, the rated engine power, the corresponding engine speed at rated power, the idle engine speed, the number of forward gears and the corresponding gear ratios. The gear shifts are determined with the commonly available information on a vehicle. The Heinz Steven tool, and the JRC (Joint Research Centre) version, based on the same legal text of the WLTP, uses the maximal torque curve of the individual vehicle. The torque curve is not part of the test procedure, and

4 therefore not properly established. In that case, however, the intermediate velocities, in particular, between 6 and 9 km/h are better covered than in the generated gear shift points in the generator. In the absence of torque information, a flat torque curve is assumed. The CADC gear shift points are based on the large European study from around 24 of normal on-road driving. This study led to many subsequent publications by M. Andre, R. Joumard and collaborators, however, the actual CADC gear shift tool seems never published. Some dependencies of gear shifts at a fixed velocities is observed with engine power and vehicle weight. The gear shift are at fixed velocities but fall in a number of vehicle classes. The NEDC gear-shift points are the basis of eco-driving, as a reverse action to adapt driving to the test optimization. For vehicles type-approved under the NEDC the vehicle is optimized for such gear shift points, and drivers are advised to use the gear shift at fixed velocities for the lowest fuel consumption. However, in some cases the vehicle has too little reserved engine power to meet real-world, or random cycle, acceleration demands. In real-world it is observed that drivers delay a gear shift when engaged in a prolonged acceleration. The Dynamic NEDC gear shifts mimic the behavior by pushing the velocity of the gear shift up with the magnitude of the acceleration prior to the intended NEDC gear shift. A graphical visualization of the generated speed profile and the gear shift advice corresponding to different gear shift strategies will be available in the plot area of the tool, by selecting the dedicated tab. F. Cycle characteristics calculation There are number of studies that focus on driving characteristics [14], [15]. These driving characteristics can be implicitly taken into consideration when generated a new driving cycle (limited number of characteristics considered to be the most representative for driving behavior) or they can be calculated after the cycle is generated, to give some more information about the cycle (extended number of characteristics). In this study, we consider a limited number of driving characteristics when generating the cycle (average speed in urban road type, minimum and maximum cycle length in seconds) and we calculate some additional cycle characteristics after the cycle was generated. For the generated speed profile a set of cycle characteristics are calculated: cycle distance, cycle time, average speed, maximum speed, road type percentage, cycle work, number of stops and other representative characteristics. These values are automatically calculated per road type as well as for the total cycle. To summarize, first a random road type profile is generated from the three different road types, in the specific order urban, rural and motorway, with random generated percentages of the total distance. Then a second random generator function is used to generate the vehicle acceleration corresponding to the already generated road type profile. Last, the speed profile is derived from the random generated acceleration profile. Also, a gear shift advice and the most important cycle characteristics are calculated. For detailed description of the road-type definition from velocities and for the velocity and acceleration cluster algorithm used in the data processing, please refer to [1]. III. WLTP RANDOM CYCLE GENERATOR TOOL Using the previously described methodology, the WLTP Random Cycle Generator Tool was developed, and it is presented in Fig. 3. This section gives a brief description of the WLTP Random Cycle Generator Tool functionality. There is a predefined distance for the generated cycles, and the order of the road types is also predefined: urban, rural and motorway, with a forced stop at the end. The percentage of the distance that will be driven in each of the above mentioned road-types is randomly generated inside the tool. There are two options available to the user: to impose vehicle limitations to the generated cycle and to generate a gear shift advice in order to follow the speed profile. When selecting this options, representative parameter values have to be filled in by the user. There is also the possibility to save vehicle configurations and to load them at a future run of the tool. After the cycle is generated, the speed profile will be plotted in the lower part of the tool, like showed in Fig. 3. Fig. 3. WLTP Random Cycle Generator Tool speed results A menu bar gives the user the option to navigate between the windows representing the speed profile, and the windows representing the four different gear advices. The calculated cycle characteristics are also available to the user, in the upper right part of the tool, like it can be seen in Fig. 4. IV. RESULTS AND VALIDATION In order to validate the stochastic approach, a number of random cycles were generated and the results were compared with the WLTP data base relative to road type, velocity and acceleration distribution.

5 6 4 2 Acceleration [m/s 2 ] Speed [m/s] Fig. 6. RCG tool VA matrix Fig. 4. WLTP Random Cycle Generator Tool gear shift results Fig. 5 shows the Velocity-Acceleration (VA) matrix of the complete WLTP data base. This contains more than 1232 hours of measurement data. 6 4 The average time share of each road type is presented in Fig. 8. Here it can be seen that, even if the distance percentage of each road type is similar, there is a difference in average time spend in each road type. This difference is given by the average driving speed per road type: low average speed in urban driving will result in high time share in this road type, while high average speed in motorway driving will result in low time share of the motorway road type. Acceleration [m/s 2 ] Speed [m/s] Share of road type percentage [%] 4 2 Urban Road type percentage [%] 1 5 Rural Road type percentage [%] 1 Motorway Fig. 5. WLTP VA matrix Fig. 6 shows the VA matrix of the tool generated data. This is based on a number of 1 cycles each with a length of 27 seconds, totaling 75 hours of stochastic generated data. Based on the similarity between the two VA matrices, it can be stated that the distribution of the generated acceleration per speed bin is representative for the real word driven data. Using the validated WLTP Random Cycle Generator tool, a large number of representative speed profile were generated, with the specific urban-rural-motorway road type order and share percentage of approximatively 34%-33%-33%. Because there is an allowed flexibility of these percentages with a ±5% range for each road type, the resulting percentage distribution per road-type is presented in Fig. 7. These results were obtained by generating 1 cycles, 35 km long Road type percentage [%] Fig. 7. Distribution of road type percentages This paper presents the development of a methodology and a software tool for generating random cycles that can be driven by a light duty vehicle on a chassis dynamometer. These randomly generated cycles are generated for the purpose of assessing criteria pollutant emission control for the type approval of light-duty vehicles. The generated random cycles are based on the WLTP database, thus enhancing the diversity of the generated cycles and making them representative for the real world driving behavior.

6 speeds. Moreover, vehicle limitations can be imposed upon the generated acceleration profile, assuring the drivability for the vehicle under test. A Beta version of the WLTP Random Cycle Generator Tool is freely available for download [16] Fig. 8. Average time share per road type V. CONCLUSIONS The RDE requirements for the European legislation are designed to be an effective means to achieve the emission reduction targets, in particular for NOx emission, without generating a large development and test burden for the automotive industry. Hence the RDE trip must reflect the European situation, such that the average emissions are reduced. Some safety margins are built-in to ensure that local emissions, e.g. urban or motorway emission separately, responsible for local high NO2 concentrations, are within bounds. However, as stated, the focus is on average emissions across Europe. In this paper a Random Cycle Generation methodology based on two Markov chains and real world measured data is developed, with the purpose of testing light duty vehicles in a simulation environment or on a test-bench. The WLTP database was used as input data. This database was created with the help of a number of European countries that collaborated and provided real world driving measurements. Consequently cycles that contain typical accelerations per velocity and road types are generated, such that these cycles are representative to real driving behavior. The presented methodology using the two-stage approach was successfully implemented in a WLTP Random Cycle Generator Tool. The speed acceleration results produced by this tool show characteristics similar to the input data; based on the similarity between the two VA matrices it can be stated that the distribution of the generated acceleration per speed bin is representative for the real word driven data. Furthermore it can be seen that the tool produces relatively steady state driving within speed-bands and shows high dynamics at lower REFERENCES [1] T. Johnson, "Review of Vehicular Emissions Trends," SAE Int. J. Engines 8(3): , 215, doi:1.4271/ [2] COMMISSION REGULATION (EU) No 459/212 of 29 May 212, [3] H.Y. Tong, W.T. Hung, C.S. Cheung, Development of a driving cycle for Hong Kong, Atmospheric Environment 33, , [4] Mario G. Perhinschi, Christopher Marlowe, Sergio Tamayo, Jun Tu & W. Scott Wayne, Evolutionary Algorithm for Vehicle Driving Cycle Generation, Journal of the Air & Waste Management Association, 211, 61:9, , DOI: 1.18/ [5] D. Karlsson and F. Tillman, Drive Cycle Prediction Using Traffic Simulation Tool, Master's Thesis, Gothenburg, Sweden 213. [6] S. Shahidinejad, E. Bibeau and S. Filizadeh, Statistical development of a duty cycle for plug-in vehicles in a North American urban setting using fleet information, IEEE Transactions on Vehicular Technology: , 21, doi:1.119/tvt [7] T-K. Lee and Z.S. Filipi, Synthesis of real-world driving cycles using stochastic process and statistical methodology, Int. J. Vehicle Design, 57(1):17-36, 211. [8] Q. Gong, S. Midlam-Mohler, V. Marano, and G. Rizzoni, An iterative markov chain approach for generating vehicle driving cycles, SAE Int. J. Engines, 4(1): , 211. [9] G. Souffran, L. Miegeville, P. Guerin, Simulation of real-world vehicle missions using stochastic Markov model for optimal design purposes, IEEE Vehicle Power and Propulsion Conference, 211, doi:1.119/vppc [1] A.E. Balau, D.G. Kooijman, I.V. Rodarte, N. Ligterink, Stochastic Real-World Drive Cycle Generation based on a Two Stage Markov Chain Approach, SAE Int. J. Mater. Manf. 8(2):39-397, 215, doi:1.4271/ [11] J. Krumm, A Markov Model for Driver Turn Prediction, SAE 28 World Congress, April 14-17, 28, Detroit, MI USA. [12] R. Feldman, C. Valdez-Flores, Applied Probability and Stochastic Processes, Springer ISBN: [13] W. Ching, M. Ng, Markov Chains: Models, algorithms and applications, International Series in Operations Research and Management Science. 26. ISBN: [14] E. Tazelaar, J. Bruinsma, B. Veenhuizen, P. van den Bosch, Driving cycle characterization and generation, for design and control of fuel cell buses, World Electric Vehicle Journal(3):1-8, 29. [15] M. André, The ARTEMIS European driving cycles for measuring car pollutant emissions, Science of the Total Environment( ):73-84, 24, doi:1.116/j.scitotenv [16] Beta version of the WLTP Random Cycle Generator Tool,