Development of Simulated Driving Cycles: Case study of the Toronto Waterfront Area

Driving cycles are an important input for state-of-the-art vehicle emission models. Development of a driving cycle requires second-by-second vehicle speed for a representative set of vehicles. Current practice is to use standard driving cycles, like HWFET or SFTP-US06, however, these driving cycles reflect small samples of vehicles, do not reflect varying conditions by time of day, by vehicle type and cannot reflect or forecast changes in traffic conditions. This paper introduces a method to develop representative driving cycles using simulated data from a calibrated microscopic traffic simulation model in the Toronto Waterfront Area. The simulation method addresses some of the limitations with standard driving cycles. The simulation model is calibrated to reflect road counts, link speeds, and accelerations using a multi-objective genetic algorithm. The simulation is validated by comparing simulated vs. observed passenger car freeway cycles. The simulation method is applied to develop AM peak hour freeway driving cycles for light, medium and heavy duty trucks. The demonstration reveals differences in speed, acceleration, and driver aggressiveness between driving cycles for different vehicle types. These driving cycles are compared against a range of available driving cycles, showing different traffic conditions and driving behaviors, and suggesting a need for city-specific driving cycles. Emissions from the simulated driving cycles are also compared with EPA’s Highway Fuel Economy Driving Schedule (HWFET) showing higher emission factors for the Toronto Waterfront freeway cycles. TRB 2013 Annual Meeting Paper revised from original submittal. G. Amirjamshidi, M. J. Roorda 3 INTRODUCTION Vehicle speed-acceleration profiles are an important input for state-of-the-art emission models. Most emission models to date have used average speed to estimate the emissions from on-road vehicles, using emissions factors. However, the emission factor method ignores fluctuations in speed around the average and acceleration rates, which are key indicators of emissions in addition to the average speed (1). Ahn and Rahka (1) demonstrate that VT-micro and CMEM (which account for speed-acceleration profiles) predict different emissions than MOBILE6 (an emissions factor model). Several more recent studies improve on average speed models by building the driving cycle for a city or a region (2-6). A driving cycle is a representative speed-time profile for a study area within which a vehicle can be idling, accelerating, decelerating, or cruising. However, speed-time profiles vary across cities due to each city’s unique topography and road driving behaviour and they have been shown to vary by vehicle type, time of day and type of road (2, 4, 6-9). The objective of this paper is to develop and demonstrate a method for efficiently developing driving cycles that represent specific combinations of roadway class, time of day and vehicle attributes. There are several applications for a disaggregate set of driving cycles. Emissions and fuel consumption impacts of changing congestion patterns, peak spreading, new infrastructure, and vehicle specific driving behavior could be better addressed. The emissions benefits of new vehicle technology could be assessed more specifically for different roadway types (e.g. when and where would the greatest benefits of plug-in hybrid electric vehicles be attained?). Vehicle routing algorithms could be developed that include congestion-sensitive fuel consumption and emissions in the objective function, for example, to service congested areas with lower emitting vehicles during off-peak hours. Two categories of driving cycles can be found in the literature. Synthesized driving cycles are built by combining different phases of constant acceleration and speed. However in such driving cycles, the transition between the different phases is unrealistic (8, 10). Examples include the European cycle and the Japanese cycle (11). Real world driving cycles are developed by recording speed-acceleration profiles while driving on the real world roadway network (often chasing a randomly selected vehicle). Examples include FTP-75 in the US, and driving cycles for Pune (8) , and Hong Kong (2). This paper introduces a method for developing representative simulated driving cycles using traffic simulation. Simulated driving cycles use simulated data from a calibrated microscopic traffic simulation for cycle development. Use of a microscopic traffic simulation allows the researcher to follow all vehicles in a network, as opposed to only a small sample of probe vehicles which can realistically be deployed to develop real world driving cycles. It also allows for analysis of the changes to the driving cycle as a result of future infrastructure or technology changes. The next section provides a review of recent studies that assess driving cycles, and provides background into the methods that are available. The following section describes the simulation based method that is used to develop AM Peak hour freeway driving cycles for the Toronto Waterfront Area for light, medium and heavy duty trucks. The resulting driving cycles are presented in the following section and compared to other available international driving cycles, to identify differences between city-specific driving cycles. Finally, the emissions that are generated for the simulated driving cycles are compared to those generated with EPA’s Highway Fuel Economy Driving Schedule (HWFET) driving cycle, to determine the extent to which emissions are sensitive to driving cycles. BACKGROUND Driving cycles have been developed for various cities and vehicle types. In general, a driving cycle is made up of micro-trips, where each micro-trip is defined as the trip made between two idling periods. A driving cycle is usually for a 10-30 min interval. This interval is long enough that there are enough micro-trips in the cycle to reflect the real world driving behaviour, but no longer than necessary so as to remain practical and reduce data collection costs (2, 6). TRB 2013 Annual Meeting Paper revised from original submittal. G. Amirjamshidi, M. J. Roorda 4 Development of a driving cycle generally involves three steps: test route selection, data collection, and cycle construction. Table 1 presents a summary of studies that develop real world driving cycles in the literature, including the study objective, the data collection method and the assessment measures that are used to evaluate the representativeness of the selected driving cycle. TABLE 1 Development Methods and Objectives of Selected Driving Cycles of Light Duty Vehicles Study Location Main Study Objective Data Collection Method Assessment Measures Montazeri and Naghizadeh (2007) Tehran (Iran) Comparing driving in Tehran vs. the FTP driving cycle Auxiliary Wheel, and photo electronic sensor Average Speed; %Time Idling Kamble et al. (2009) City of Pune, (India) Cycle Construction and comparison to IDC and ECE-15 + EUDC Car chasing Average Speed; %Time Accelerating, Decelerating, Cruising, Idling Coelho et al. (2009) Mol (Belgium) Comparing driving cycles and emission of light duty gasoline vs. diesel vehicles Portable Emission Measurement System (PEMS) and GPS device 14 modes of Vehicle Specific Power (VSP) Hung et al. (2007) Hong Kong Introducing a practical method for developing real world driving cycles On-board measurement, and car chasing Average speed, Running speed, Acc, Dec, Micro-trip duration; %Time Accelerating, Decelerating, Cruising, Idling, Creeping; Root mean square acceleration, and Positive kinetic energy Wang et al. (2008) 11 cities in China Development and comparison of driving cycles for 11 cities Car chasing Same as Hung et al. excluding Average micro-trip duration; % Time creeping Saleh et al. (2009) Edinburgh (UK) Development and comparison of rural vs. urban driving cycles Performance Box (PB) Same as Hung et al. except: % Time creeping Tzirakis et al. (2006) Athens (Greece) Development and comparison of Athens’ urban driving vs. the European driving cycle Car chasing Previously developed software accounting for topographical and other characteristics Yu et al. (2010) Houston (Texas) Use of a genetic algorithmbased approach for developing driving cycles comparing four assessment measures PEMS and GPS device • Average speed, Running speed, Acc, Dec; Maximum speed; %Time Accelerating, Decelerating, Idling, Constant speed; • 17 bins of Vehicle Specific Power (VSP) • Average fuel consumption rate for each VSP bin based on MOVES • Product of VSP and fuel consumption Test route selection involves selecting the route on which data are to be collected. The intention is to select routes that exhibit vehicle motion that is representative of typical driving conditions of the full population of vehicles. The ability of a test route to be fully representative is, of course, limited. Test TRB 2013 Annual Meeting Paper revised from original submittal. G. Amirjamshidi, M. J. Roorda 5 route selection is not necessary for simulated driving cycles, since data can be collected on all routes within a desired roadway classification. The data collection step generally involves the collection of the speed of a sample of vehicles at frequent time intervals (usually on a second-by-second basis). The data collection method most commonly reported in the literature is to make on-board measurements on probe vehicles and/or car chasing (Table 1). Collection of real world data to develop driving cycles for different vehicles and road types, for a large enough representative sample of vehicles would either be too costly or biased (if data were collected on a day with unusual congestion patterns). For this reason, we propose the use of simulated traffic data, which can be collected for all vehicles under consistent and calibrated traffic conditions. Using data from multiple simulation replications also accounts for stochastic variations in traffic conditions, allowing a driving cycle to be more representative. However, for simulation to be considered suitable for driving cycle construction, calibration must be undertaken for the elements of the cycle that are important, including traffic counts, vehicle speeds and variation in acceleration. The third step, cycle construction, consists of the following: 1) Define the set of assessment measures used to describe a driving cycle; 2) Calculate the assessment measures for the collected data (called target statistics); 3) Develop a candidate driving cycle from the pool of micro-trips available (called candidate cycle); 4) Calculate the same assessment measures for the candidate cycle (called test statistics); 5) Identify the candidate cycle whose test statistics are closest to the target statistics. All except one study in the literature have used random selection of micro-trips as the method for producing a candidate cycle. Yu et al. (6) used a semi-genetic algorithm based method, which calculates the assessment measures for each micro-trip, sorts the micro-trips in ascending order based on their measures, selects a micro-trip from the top of the list, calculates the lower and upper limit for the number of required micro-trips, and tests all combinations of micro-trips within the selected pool. Although this method improves slightly upon the random selection method, the large computational time is prohibitive. More importantly, the authors point out that their method could converge at a local minimum. Consequently, in this research, the random selection method was chosen. The other difference between studies is the type and number of assessment measures used (as can be seen in Table 1). Driving activity measures and the Vehicle Specific Power (VSP) method have both been widely used in the literature. Activity measures refer to the use of statistics like speed and acceleration (2, 3, 9, 12). The VSP method, which is a bin based method, focuses on instantaneous power per unit of a vehicle and is a nonlinear function of instantaneous speed, instantaneous acceleration and road grade (5, 6). In Yu et al. (6) the VSP method produced 2% less CO2 prediction error compared to the driving activity method. However it should be noted that they used nine of the widely used driving activity measures, where most studies using driving activity measures use more measurements (2-4). Using more driving activity measures would likely reduce the difference in CO2 prediction error between the VSP and the driving activity method. Also Yu et al. (6) used data points for all road types in their comparison, which would affect their results and comparison in selecting the best method for cycle development. Hence in this research the driving activity measures presented by Hung et al. (2) are used and are discussed in detail in the method section. METHOD The only recent driving cycles developed for the Toronto area were synthesized using CALMOB6 (10) reflecting average speeds from a travel demand model using the Transportation Tomorrow Survey (TTS) (13). To the best of our knowledge no real world driving cycle has been developed for the Toronto area. In this research, the concept of developing a simulated driving cycle is introduced and the method is demonstrated to develop freeway driving cycles for passenger cars, light duty trucks (LDT), medium duty TRB 2013 Annual Meeting Paper revised from original submittal. G. Amirjamshidi, M. J. Roorda 6 trucks (MDT), and heavy duty trucks (HDT) in the Toronto Waterfront Area. The method can easily be extended to obtain a full suite of driving cycles by time of day, by vehicle type and by roadway type. Traffic Simulation Model Study Area The traffic simulation is developed for the Toronto Waterfront Area (Figure 1) which is located south of Dundas Street, west of Woodbine Avenue, and east of Parkside Drive. This area consists of the central business district of Toronto and inner urban areas to the east and west. The network includes arterial, collector and local roads and two freeways, the Don Valley Parkway (DVP) and the Gardiner Expressway, that play an important role in transporting goods and people to and from the downtown. The network was originally coded in Paramics V5 (the microscopic traffic simulation model) in a project conducted for the Toronto Waterfront Revitalization Corporation (14). Within that project, efforts were invested into building the correct geometry, defining the roadway attributes (speeds, and land configurations) and coding signal timing. For signalized intersections, actuation algorithms were developed to best represent the SCOOT traffic signal control system in the Waterfront area. Detailed information about the steps taken is available in (14). The model had been calibrated for 2001 traffic conditions and vehicle demand; therefore, significant additional calibration was required for this research to update the model for 2009 traffic conditions. The final network in Paramics V6.8 (15) consists of 4012 roadway links, 1841 nodes, 44 internal zones, 35 external gateways, and 227 signals, and approximately 26 km of freeway. FIGURE 1 Toronto Waterfront Area Network. Demand inputs were generated using a multiclass generalized cost static user equilibrium assignment (in the EMME modeling software) for the Greater Toronto and Hamilton Area for light, medium and heavy trucks and passenger cars. Passenger vehicle demand was derived from the household travel survey in Toronto, and truck demand was developed using a three stage truck trip model based on data from a shipper based survey of truck demand and truck roadside interview data. The demand was calibrated at the regional level to reflect traffic counts at cordons across the region, and was further calibrated for the Toronto Waterfront Area using OD matrix updating. The steps for this stage and the data sources used to develop the model are described in detail in Roorda et al. (16). Calibration For calibrating the traffic simulation model for the AM peak hour, a multi-objective optimization, requiring optimization on more than one criterion, was undertaken using a genetic algorithm for parameter estimation. TRB 2013 Annual Meeting Paper revised from original submittal. G. Amirjamshidi, M. J. Roorda 7 There are two methods for dealing with a multiobjective problem: the Scalar method, and the Pareto method (17, 18). The scalar method was selected because it is relatively easy to implement and computationally efficient. In this method all objective functions are combined to form a single value using linear or non-linear functions. Since speed and acceleration profiles are essential in estimating emissions, the objective function for this study (Equation 1) was chosen so that speed and variation in acceleration would also be calibrated. MSE = � � 1 n1 ∑ (SimCounti−ObsCounti)2 n1 i=1 AvgCount + � 1 n2 ∑ (SimSpeedi−ObsSpeedi)2 n2 j=1 AvgSpeed + � 1 n2 ∑ (SimStd(acc)i−ObsStd(acc)i)2 n2 j=1 AvgStd(acc) �