View metadata, citation and similar papers at core.ac.uk brought to you by CORE

provided by Elsevier - Publisher Connector

Scientia Iranica A (2013) 20 (2), 286–293

Sharif University of Scientia Iranica Transactions A: Civil www.sciencedirect.com

Research note Tehran driving cycle development using the k-means clustering method

A. Fotouhi ∗, M. Montazeri-Gh Systems Simulation and Control Laboratory, School of , Iran University of Science and Technology, Tehran, P.O. Box 16846-13114, Iran

Received 30 November 2011; revised 7 October 2012; accepted 29 October 2012

KEYWORDS Abstract This paper describes the development of a car driving cycle for the city of Tehran and its suburbs Driving cycle; using a new approach based on driving data clustering. In this study, driving data gathering is performed k-means clustering; under real traffic conditions using Advanced Vehicle Location (AVL) devices installed on private cars. The Vehicle; recorded driving data is then analyzed, based on ‘‘micro-trip’’ definition. Two driving features including Fuel consumption; ‘‘average speed’’ and ‘‘idle time percentage’’ are calculated for all micro-trips. The micro-trips are then Exhaust emissions. clustered into four groups in driving feature space using the k-means clustering method. For development of the driving cycle, the nearest micro-trips to the cluster centers are selected as representative micro- trips. The new method for driving cycle development needs less computation compared to the SAPM method. In addition, it benefits the capability of the k-means clustering method for traffic condition grouping. The developed driving cycle contains a 1533 s speed time series, with an average speed of 33.83 km/h and a distance of 14.41 km. Finally, the characteristics of the developed driving cycle are compared with some other light vehicle driving cycles used in other countries, including FTP-75, ECE, EUDC and J10-15 Mode. © 2013 Sharif University of Technology. Production and hosting by Elsevier B.V. under CC BY-NC-ND license.

1. Introduction the NEDC used in Europe and the J10-15 used in Japan, are sam- ples of legislative driving cycles. Non-legislative driving cycles, Air pollution is a major problem for big cities, where vehicles such as the Hong Kong driving cycle [1], are used in research for are one of the most important air pollution sources. On the other energy conservation and pollution evaluation. hand, car fuel consumption has a great share in energy usage. There are two approaches for developing a driving cycle. Vehicle fuel consumption and exhaust emissions are affected In the first method, a driving cycle is composed of various by driving patterns formed under different traffic conditions. driving modes of constant acceleration, deceleration and speed, ‘‘Driving Cycles’’ are speed-time profiles which represent and is referred to as ‘‘modal’’ or ‘‘polygonal’’, such as NEDC driving patterns in a region or city. They are used to simulate and ECE driving cycles [2–4]. In the second method, a driving driving conditions on a laboratory chassis dynamometer for cycle is derived from actual driving data and is referred to evaluation of fuel consumption and exhaust emissions. as a ‘‘real world cycle’’, such as FTP-75. The real world cycles There are two major categories of driving cycle including are more dynamic, reflecting more rapid acceleration and legislative and non-legislative. According to legislative driving deceleration patterns experienced during driving conditions. cycles, exhaust emission specifications are imposed by govern- This more dynamic driving in real world conditions leads to ments for car emission certification. The FTP-75 used in the USA, higher emissions compared to those on the modal test cycles. As driving patterns vary from city to city and from one area ∗ Corresponding author. to another, the available driving cycles obtained for certain E-mail address: [email protected] (A. Fotouhi). cities or countries are not usually applicable for other cities and under responsibility of Sharif University of Technology. countries. Therefore, much research work is targeted towards developing driving cycles for their own city or region [5–8]. Development of a car driving cycle for the city of Tehran was also investigated in previous studies, based on statistical analysis of the driving data, containing the Speed-Acceleration

1026-3098 © 2013 Sharif University of Technology. Production and hosting by Elsevier B.V. Open access under CC BY-NC-ND license. http://dx.doi.org/10.1016/j.scient.2013.04.001 A. Fotouhi, M. Montazeri-Gh / Scientia Iranica, Transactions A: 20 (2013) 286–293 287

Figure 1: The flowchart of driving cycle development approach.

Probability Matrix (SAPM) and the chi square criterion [9,10]. cycle. Another advantage of the new approach is that it is not However, the SAPM method requires large computations for sensitive to types and number of driving features. The approach selection of the representative micro-trips. In that method, the is simply applicable using any other types or any number SAPMs of all micro-trips should be constructed separately and, of driving features. Besides the differences between the new then, they must be compared to the total SAPM in order to rank driving cycle development approach and the old one [10], this them [10]. study aims to update the Tehran driving cycle, due to changes In this paper, the development of a car driving cycle for in traffic flow and driving patterns over the years. In addition, the city of Tehran and its suburbs is presented using a new the new driving cycle is developed for the city of Tehran and approach based on driving data clustering. In the new approach, its suburbs, whereas, in the previous study, data collection had the process of micro-trip selection is performed with less been performed only in the city itself and not in the suburbs. computation. In the proposed method, the clustering of micro- The structure of the paper is as follows. The driving trips is done instead of constructing a SAPM. The nearest micro- cycle development approach is described in Section 2, briefly. trips to cluster centers are then selected to form the final driving In Section 3, data gathering and micro-trip definition are 288 A. Fotouhi, M. Montazeri-Gh / Scientia Iranica, Transactions A: Civil Engineering 20 (2013) 286–293 presented. The k-means clustering method is described in Section4. Section5 contains feature extraction, micro-trips clustering and traffic condition definitions. Finally, in Section 6, the newly developed Tehran driving cycle is presented and analyzed.

2. Driving cycle development approach

In this paper, driving cycle development is performed using a new approach, based on driving data clustering. In this approach, driving data collection is firstly done under real traffic condition. Driving data is then divided into small parts called ‘‘micro-trips’’. Subsequently, driving feature extraction is carried out in order to characterize micro-trips. Driving Figure 2: Advanced Vehicle Location (AVL) system. features are calculated for each micro-trip and each micro-trip is mapped into a feature space as a point. After that, micro- trips are clustered into groups in the feature space using the k-means clustering method. Each cluster is called a ‘‘traffic condition’’ and one sub-cycle is made for each traffic condition. The final driving cycle consists of the separate sub-cycles of traffic conditions. The driving cycle of each traffic condition consists of some representative micro-trips from that traffic condition. In this step, a procedure is required to select the representative micro-trips. In the proposed approach of this study, the nearest micro-trips to cluster centers are selected as representative micro-trips of each cluster. The flowchart of the Figure 3: Sample micro-trips. driving cycle development approach is depicted in Figure 1. The steps of the procedure are described in the following sections. during six months under real traffic conditions. In order to The share of each traffic condition in the final driving cycle analyze the driving data, a partitioning approach is proposed in is proportionate to the length of micro-trips of that cluster in this study, based on the definition of ‘‘micro-trips’’. A Micro-trip all driving data. In other words, to obtain the duration of each is an excursion between two successive time points at which traffic condition in the final driving cycle, the time proportion of the vehicle is stopped [11]. This part of the motion consists of each cluster in the whole recorded data is used. It is formulated acceleration, cruise and deceleration modes. Figure 3 depicts as follows: a sample driving time series containing 6 micro-trips. This

ni partitioning approach is required for driving feature extraction, tdriving cycle  ti = ti,j, (1) and clustering of the driving data. t Overall j=1 where: 4. k-means clustering method

ti is duration of cluster number i (i = 1 to number of clusters) in the final driving cycle, In this study, the k-means clustering method is utilized for micro-trips clustering. In this section, the k-means clustering tdriving cycle is the duration of the final driving cycle, algorithm is explained briefly. Clustering in N-dimensional toverall is the duration of all recorded data, Euclidean space, RN , is the process of partitioning a given set ti,j is the time of micro-trip number j in cluster number i, and of n points into a number, say K, of groups (or, clusters), based ni is the total number of micro-trips in cluster number i. on some similarity/dissimilarity metric [12]. Let the set of n There should be agreement between the representative points {x1, x2,..., xn} be represented by set S, and K clusters degree of a cycle and the low cost test procedure on the be represented by C1, C2,..., CK . Then: dynamometer, while the former implies long duration and the latter requires a short duration of driving cycle. It is more Ci ̸= ∅ for i = 1,..., K appropriate to develop a cycle with adaptation to reference Ci ∩ Cj = ∅ for i = 1,..., K, j = 1,..., K and i ̸= j cycle characteristics and with short duration suitable for dynamometer tests. Considering existing driving cycles, it and: seems that a cycle with duration of about 30 min may well K represent a driving situation in a city as a reference cycle, and  C = S (2) about 15 min as a test cycle. i . i=1

3. Data collection and micro-trip definition One of the most widely used clustering techniques available in the literature is the k-means algorithm [13,14]. The k- In this study, an Advanced Vehicle Location (AVL) device is means algorithm attempts to solve the clustering problem by employed for data collection. The AVL device, shown in Figure 2, optimizing a given metric. The steps of the k-means algorithm is an advanced device for vehicle tracking and monitoring, are described here, briefly [12]: based on GPS technology. Driving data collection is performed Step 1: Choose K initial cluster centers, z1, z2,..., zK , by private cars moving in different places of the city of Tehran randomly from n points, {x1, x2,..., xn}. A. Fotouhi, M. Montazeri-Gh / Scientia Iranica, Transactions A: Civil Engineering 20 (2013) 286–293 289

Step 2: Assign point xi, i = 1, 2,..., n to cluster Cj, j ∈ {1, 2,..., K}, if:

∥xi − zj∥ < ∥xi − zp∥, p = 1, 2,..., K, and j ̸= p. (3) Ties are resolved arbitrarily. ∗ ∗ ∗ Step 3: Compute new cluster centers, z1 , z2 ,..., zK , as follows: 1 ∗ =  = zi xj, i 1, 2,..., K, (4) ni xj∈Ci where ni is the number of elements belonging to cluster Ci. ∗ = = Step 4: If zi zi, i 1, 2,..., K, then terminate. Otherwise, continue from Step 2. Note that in cases where the process does not terminate at Step 4, normally, then it is executed for the maximum number of iterations.

5. Feature extraction, micro-trip clustering and traffic condition definition Figure 4: Scatter plot of the micro-trips in the 2-D feature space. As mentioned earlier, the proposed approach of this study for driving cycle development is based on micro-trip clustering. For this purpose, driving features should be extracted first. Many driving features have been defined for driving data [15–18]. In this study, two common driving features, including ‘‘average speed’’ and ‘‘idle time percentage’’ have been used. These two parameters were chosen as they have the greatest effects on emissions [19]. It should be noted that the proposed approach in this study is applicable for any type and number of driving features. The two mentioned driving features are formulated as follows:

1- Average speed (Vmean): Average speed of a micro-trip (n is the length of the micro-trip in sec, vi is the velocity value in seconds i).

1 n Vmean = vi. (5) n i=1 2- Idle time percentage: Time percentage in a micro-trip, where speed is zero (i.e. the vehicle stops). Figure 5: Clustering of the micro-trips in the 2-D feature space. Using the two above driving features, each micro-trip can be plotted as a point in 2-dimensional feature space. The scatter 6. Tehran driving cycle development and analysis plot of all micro-trips in the feature space is presented in Figure 4. As seen in this figure, there is a reasonable relationship between the driving features, such that the more the idle time In previous section, micro-trips were clustered into four is, the less the average velocity is. For micro-trips with very traffic conditions. In this section, representative micro-trips small idle time, the average speed is high. On the other hand, are determined in order to make a driving cycle for each for micro-trips with greater idle time, the average speed is very cluster. For this purpose, the nearest micro-trips to cluster low. centers are selected as the representative micro-trips of each In Figure 5, the micro-trips are clustered into four groups cluster. The process of micro-trip selection continues to select using the k-means clustering method. Each cluster has its own enough micro-trips from all clusters. As mentioned in Section 2, characteristics and stands as a traffic condition. Consequently, the share of each traffic condition in the final driving cycle four traffic conditions are considered for Tehran driving cycle is proportionate to the length of micro-trips of that traffic as follows: condition in all driving data. Time percentage and duration of 1- Congested traffic condition: For central business district each traffic condition in the Tehran driving cycle are presented flows with very low driving speed and frequent stops which in Table 1. Considering the table, the selection of micro-trips means high idle time (depicted as cluster 1 in Figure 5). from each cluster continues until the time sharing requirement 2- Urban traffic condition: For non-free flows with moderate is met. The selected micro-trips of the four traffic conditions are and low idle time and low average speed (depicted as cluster demonstrated in Figures 6–9. Putting the sub cycles of different 2 in Figure 5). traffic conditions beside each other, the Tehran driving cycle is 3- Extra urban traffic condition: For arterial routes with relatively free flows and low idle time and moderate average constructed as presented in Figure 10. speed (depicted as cluster 3 in Figure 5). Following the construction of an initial driving cycle, it 4- Highway traffic condition: For completely free flows with should be smoothed, in order to be suitable for a dynamometer very low idle time and high average speed (depicted as test. For this purpose, a filter has been applied in order cluster 4 in Figure 5). to remove the high frequency noises which make a driving 290 A. Fotouhi, M. Montazeri-Gh / Scientia Iranica, Transactions A: Civil Engineering 20 (2013) 286–293

Table 1: Time percentage and duration of different traffic conditions in Tehran driving cycle. Time percentage Duration in driving cycle (s) in driving cycle (%)

Congested 11.35 174 Urban 8.48 130 Extra urban 16.57 254 Highways 63.6 975 Total 100 1533

Figure 8: Extra urban driving cycle.

Figure 6: Congested driving cycle.

Figure 9: Highway driving cycle.

Figure 7: Urban driving cycle. Figure 10: Tehran driving cycle containing four traffic conditions. cycle difficult to follow. The filter is defined by the following The initial and smoothed driving cycles are illustrated in formula [11]: Figure 11. As shown in the figure, the sharp points are smoothed using the filter. h 1   s  In order to evaluate the developed driving cycle, it is vsmoothed(t) = K · v(t + s). (6) h h compared to some other light vehicle driving cycles used in s=−h other countries, such as FTP-75, ECE, EUDC and J10-15 Mode, Function k(x) weighs the measured speed just before and after which are illustrated in Figures 12–15. Comparison between time t to be smoothed. In this study, h = 4 s, together with the characteristics of the Tehran driving cycle and other driving so-called ‘‘biweight’’ smoothing kernel, has been used [11]: cycles is presented in Table 2. The results demonstrate that the developed driving cycle for  2  h − 1 the city of Tehran contains 1533 s, with an average speed of  (1 − x2)2 (x2 < 1)  K(x) = h2 . (7) 33.83 km/h. A car passes a distance of 14.41 km over the driving  0 otherwise cycle. Compared to other driving cycles, the Tehran driving A. Fotouhi, M. Montazeri-Gh / Scientia Iranica, Transactions A: Civil Engineering 20 (2013) 286–293 291

Table 2: Comparison of Tehran driving cycle and other light vehicle cycles.

Time Distance Maximum Average Maximum Maximum Average Average Idle (s) (km) veloc- veloc- accelera- decelera- accelera- decelera- time (%) ity (km/h) ity (km/h) tion (m/s2) tion (m/s2) tion (m/s2) tion (m/s2) Tehran 1533 14.41 91.06 33.83 2.15 −1.88 0.47 −0.49 15.26 FTP-75 2477 17.67 90.72 25.67 1.47 −1.47 0.51 −0.57 38.8 ECE 195 0.99 49.71 18.06 1.05 −0.83 0.64 −0.74 33.33 ECE + EUDC 1225 10.87 119.30 31.91 1.05 −1.38 0.54 −0.78 27.67 J10-15 mode 660 4.14 69.56 22.55 0.79 −0.83 0.53 −0.59 32.58

Figure 11: Smoothed driving cycle.

Figure 14: EUDC driving cycle.

Figure 12: FTP-75 driving cycle.

Figure 15: J10-15 Mode driving cycle.

(ADVISOR) software has been utilized for the simulations. The software has been used in previous studies for estimation of vehicle fuel consumption and exhaust emissions [20]. A common vehicle in Iran, called ‘‘Samand’’, is investigated in the simulations. Specifications of the Samand vehicle are presented in Table 3. The vehicle is simulated on FTP, ECE, and EUDC driving cycles, as well as the newly developed Tehran driving cycle. The simulation results are depicted in Figures 16–19. As demonstrated in Figure 16, the FC value of the vehicle is 8.77 (L/100 km) in the Tehran driving cycle, which is less than Figure 13: ECE driving cycle. the FC value on the ECE driving cycle and more than the two others. Similar to the FC, the HC and NOx emissions of the cycle has a higher average speed and lower idle time percentage vehicle on the Tehran driving cycle are in the middle range (i.e. 15.26%). This is justified, as the developed driving cycle compared to the three other driving cycles. The CO emission has has resulted from the data collection in Tehran city and its a different pattern, in which the emission value over the Tehran suburbs, where the majority of trips occur under highway traffic driving cycle is a little more than the other driving cycles. conditions. Moreover, the influence of driving cycles on vehicle Fuel Consumption (FC) and exhaust emissions is investigated using 7. Conclusion computer simulations. Three exhaust emissions are considered, including unburned hydrocarbons (HC), carbon monoxide (CO) In this paper, a methodological approach is presented for and oxides of nitrogen (NOx). Advanced vehicle simulator driving cycle development based on micro-trips clustering. 292 A. Fotouhi, M. Montazeri-Gh / Scientia Iranica, Transactions A: Civil Engineering 20 (2013) 286–293

Table 3: Specifications of vehicle.

Vehicle parameters Parameter Value Unit

1 Total vehicle mass 1250 kg 2 Vehicle glider mass 905 kg 3 Coefficient of aerodynamic drag 0.318 Vehicle 4 Frontal area 2.1 m2 5 Fraction of vehicle weight on front axle 0.6 6 Height of vehicle center-of-gravity 0.64 m 7 Wheelbase 2.67 m 8 Peak engine power 82 kW 9 Rotational inertia of the engine 0.18 kg m2 10 Total engine/fuel system mass 262 kg 11 Fuel density 0.753 g/l 12 Lower heating value of the fuel 42.5 kJ/g Combustion engine 13 Exterior surface area of engine 0.2626 m2 14 Air/fuel ratio (stoic) on mass basis 14.5 15 Engine coolant thermostat set temperature 86 ° C 16 Average cp of engine 500 J/kg K 17 Average cp of hood and engine compartment 500 J/kg K 18 Surface area of hood/eng compt. 1.5 m2

Figure 18: CO emission on different driving cycles.

Figure 16: FC on different driving cycles.

Figure 19: NOx emission on different driving cycles.

Figure 17: HC emission on different driving cycles. of 33.83 km/h and 15.26% idle time. The car passes 14.41 km over the Tehran driving cycle. Compared to ECE, EUDC and FTP The new approach is simply applicable for any type or driving cycles, the newly developed cycle has higher average number of driving features. This proposed approach needs less speed and a lower idle time percentage, and it causes vehicle computation compared to previous methods, such as the SAPM FC and emissions in the middle range. method, and it has proper flexibility to face different types of driving data. In addition, different traffic conditions of a driving cycle are determined as a result of the micro-trip clustering References in the new approach. A driving cycle is developed for the city [1] Tong, H.Y., Hung, W.T. and Cheung, C.S. ‘‘Development of a driving cycle of Tehran and it suburbs using the proposed approach. The for Hong Kong’’, Journal of Atmospheric Environment, 33, pp. 2323–2335 developed driving cycle contains 1533 s, with an average speed (1999). A. Fotouhi, M. Montazeri-Gh / Scientia Iranica, Transactions A: Civil Engineering 20 (2013) 286–293 293

[2] Boulter, P.G. and Cox, J.A., ‘‘A review of European emission measurements [15] Ericsson, E. ‘‘Variability in urban driving patterns’’, Journal of Transporta- and models for diesel fueled buses’’, TRL Report 378, Crowthorne, (1999). tion Research, Part D, 5, pp. 337–354 (2000). [3] Elgeneman, M., Sorusbay, C. and Goktan, A. ‘‘Development of a driving [16] Ericsson, E. ‘‘Independent driving pattern factors and their influence on cycle for the prediction of pollutant emissions and fuel consumption’’, fuel-use and exhaust emission factors’’, Journal of Transportation Research, International Journal of Vehicle Design, 18(3–4), pp. 391–399 (1997). Part D, 6, pp. 325–341 (2001). [4] Kuhler, M. and Karstens, D. ‘‘Improved driving cycle for testing automotive [17] Brundell-Freij, K. and Ericsson, E. ‘‘Influence of street characteristics, exhaust emissions’’, SAE Technical Paper, Series 780650, 1978. driver category and car performance on urban driving patterns’’, Journal [5] Lin, J. and Niemeier, D.A. ‘‘Regional driving characteristics, regional driving of Transportation Research, Part D, 10, pp. 213–229 (2005). cycles’’, Journal of Transportation Research, Part D, 8(5), pp. 361–381 (2003). [18] Montazeri-Gh, M., Fotouhi, A. and Naderpour, A. ‘‘Driving patterns [6] Hung, W.T., Tong, H.Y., Lee, C.P., Ha, K. and Pao, L.Y. ‘‘Development of a clustering based on driving features analysis’’, Proceedings of the Institution practical driving cycle construction methodology: a case study in Hong of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science, Kong’’, Journal of Transportation Research, Part D, 12, pp. 115–128 (2007). 225(6), pp. 1301–1317 (2011). [7] Kamble, S.H., Mathew, T.V. and Sharma, G.K. ‘‘Development of real-world [19] Kent, J.H., Allen, G.H. and Rule, G. ‘‘A driving cycle for Sydney’’, Journal of driving cycle: case study of Pune, India’’, Journal of Transportation Research, Transportation Research, 12, pp. 147–152 (1978). Part D, 14, pp. 132–140 (2009). [20] Markel, T. and Brooker, A. ‘‘ADVISOR: a systems analysis tool for advanced [8] Nutramon, T. and Supachart, Ch. ‘‘Influence of driving cycles on exhaust vehicle modeling’’, Journal of Power Sources, 110, pp. 255–266 National emissions and fuel consumption of gasoline passenger car in Bangkok’’, Renewable Energy Laboratory, Golden, CO 80401, USA (2002). Journal of Environmental Sciences, 21, pp. 604–611 (2009). [9] Montazeri-Gh, M., Varasteh, H. and Naghizadeh, M. ‘‘Driving cycle simulation for heavy duty engine emission evaluation and testing’’, SAE, Abbas Fotouhi has a Ph.D. degree in Mechanical Engineering, and has been 2005-1-3796, 2006. carrying out research in the Systems Simulation and Control Laboratory at [10] Montazeri-Gh, M. and Naghizadeh, M. ‘‘Development of the Tehran car Iran University of Science and Technology, Tehran, Iran. His research interests driving cycle’’, International Journal of Environment and Pollution, 30(1), include: electric, hybrid electric and fuel cell vehicles, driving and traffic pp. 106–118 (2007). conditions analysis, vehicle dynamics and control. [11] Haan, P.D. and Keller, M., ‘‘Real-world driving cycles for emission measurement: ARTEMIS and Swiss cycles’’, Final Report, (2001). [12] Maulik, U. and Bandyopadhyay, S. ‘‘Genetic algorithm based clustering Morteza Montazeri-Gh established the Systems Simulation and Control technique’’, Pattern Recognition, 33(9), pp. 1455–1465 (2000). Laboratory, in 1996, at Iran University of Science and Technology, Tehran, Iran, [13] Kaufman, L. and Rousseeuw, P.J., Finding Groups in Data: An Introduction to where he is currently Professor of Mechanical Engineering. His main research Cluster Analysis, Wiley (1990). interests include: systems simulation and control, and, especially, the use of [14] Hartigan, J.A., Clustering Algorithms, Wiley, New York (1975). driving data in intelligent HEV control.