Track: Smart Mobility and Smart Cities

Abstract

A Driving Pattern (DP) describes how drivers drive in a given region. The DPs are obtained by monitoring a representative sample of vehicles traveling on roads of different characteristics and under normal conditions. A Driving Cycle (DC), on the other hand, is a speed time series that represent the DP, and it is used to assess the performance in energy consumption and tailpipe emissions of the vehicles. The construction of a DC consists in obtaining a speed time series whose characteristic parameters (CP*s) are similar to the characteristic parameters of the driving pattern (CPs). The CPs are variables derived from speed and time, such as the mean speed and mean acceleration. The methods to construct DC (Micro-Trips, Markov chains, and MWD-CP) are well-known and established. However, challenges remain with respect to the repeatability of their results, the need for a large number of measured trips, which is highly expensive and time-consuming.

This work presents an alternative and novel approach to construct artificial DC representing DPs using a few measured trips and a Deep Learning (DL) model called Generative Adversarial Networks (GAN). A GAN combines two DL architectures: the generator and the discriminator. The generator creates candidate artificial DC as similar as possible to the real trips, while the discriminator distinguishes between real and artificial DC. Therefore, a competition is triggered between the generator and discriminator, resulting in a generator capable of creating representative candidate artificial DC.  For instance, the generator looks after: i) to minimize loss functions which are the relative differences between CPs of the artificial DC and of the driving patterns, and ii) to maximize the failure of the discriminator. In addition, the resulting model can be used to obtain DCs from other cities using transfer learning techniques. Transfer learning consists of reusing a well-trained and validated model in a similar application by adjusting only the model parameters.

As a case example, we collected second by second data of speed, fuel consumption, and tailpipe pollutants, obtained monitoring for two months, a sample of 16 vehicles operating in Mexico City. A GAN was trained for constructing artificial but representative DC of 20 minutes duration with a resolution of 1 Hz. We used that GAN to construct 10000 representative DC and found that in all cases, the DCs exhibit CP*s within the 10% of the relative difference with respect to the corresponding 19 CPs that describe the DP, fuel consumption, and tailpipe emissions. According to the literature, 10% of the relative difference is an acceptable criterion for this purpose. Similar results were observed applying transfer learning to data from Bogota, Colombia.

The proposed artificial trip generator aims to alleviate the data collection costs to obtain DC, fuel consumption, and tailpipe emissions. Ultimately, we introduced new emerging technologies based on artificial intelligence to tackle an existing challenge in the field of intelligent transportation and smart mobility.