Threshold-guided multi-objective Generative Adversarial Networks for constructing artificial yet representative driving cycles

A Driving Cycle (DC) is a speed-time series that represents how people drive in a region. It is used to assess the vehicle's energy and environmental performance. Constructing representative DCs has been widely studied with well-known methods. However, challenges remain regarding the reproducibility of results and the need for a large set of sampled trips, which can be costly and time-consuming. This work presents a novel approach to constructing DCs using Generative Adversarial Networks (GANs). A multi-objective training approach incorporates model-based losses represented by discriminators to capture time-series trends of speed, acceleration, and jerk, and function-based losses to ensure that the generator respects physical constraints. These losses are activated sequentially according to a convergence measure that utilizes the Characteristic Parameters (CPs) of the generated DCs for the latest activated loss. A new loss is activated if the convergence measure is lower than a threshold. An early stopping mechanism can be triggered when the threshold condition is met at the last loss. The DC generator was validated comparing the artificial trips with real trips from four cities. This comparison involved computing the Relative Difference (RD) between the CPs. The results indicate that the approach can achieve less than a 10% RD, thus providing superior results than a classical approach, a vanilla GAN, and other variations. Additionally, the threshold-based loss activation led to a reduction in training time. Transfer Learning (TL) was employed to construct DCs for other regions. This approach reduces the need for extensive data collection and ensures the reproducibility of artificial DCs.