Capturing local vehicle behavior in city-specific typical driving cycles using machine learning: Case of Bogota, Colombia

This work presents a novel methodology based on machine learning for developing city-representative typical driving cycles (TDCs) of light-duty vehicles. The methodology incudes the combination of principal component analysis to reduce dimensionality of the parameters space, k-means clustering with metaheuristics for micro-trips clusterization, and Markov chains to connect representative micro-trips into fully computed TDCs. Existing on-road datasets from the U.S. cities of El Paso, San Antonio, and Houston/Galveston were used to validate the methodology by computing three corresponding local TDCs. The computed TDCs were assessed in their representativeness of each city's dataset and compared to standard TDCs. Representativeness assessments show that the proposed methodology is robust, and that the various metaheuristics applied do not have a significant effect on the average error between the computed TDCs and the datasets. Comparisons demonstrate that the computed city-specific TDCs present much smaller mean error values than the standard TDCs used for regulatory purposes. Finally, the validated methodology was applied to on-road test data in Bogota metro areas, producing a local ‘Bogota TDC’ with average mean errors lower than 5 % with respect to the dataset. The typical driving emissions and fuel consumption obtained from measurements on a chassis dynamometer using the Bogota TDC are also reported. Notably, the dynamometer tests with the Bogota TDC reveal ∼25 % higher fuel consumption compared to standard TDCs. These findings highlight the importance of developing representative local TDCs to better estimate vehicles’ near-Real World behavior, as well as fuel economy and emission measurements for regulatory compliance with environmental standards.