Large-scale optimization of electric vehicles using graphics processing units

Abstract

Large-scale design optimizations of electric vehicles have been limited in the past due to a lack of computational power. In order to fully utilize the potential of EV technology, novel optimization methods need to be developed. This dissertation investigates new opportunities for EV system optimization evolving from recent advancements in parallel processing hardware. The power train and control system of a fuel cell hybrid electric vehicle are optimized on the parallel architecture of graphics processing units. A two-level optimization methodology is developed specifically for peak performance on GPU architectures. Computational speedup factors of more than 2,000x and 70,000x are achieved over a sequential C/C++ implementation and the Matlab/Simulink environment, respectively. The significant gain in computation speed enables incorporation of expensive lifetime effects, such as battery degradation, as well as parameter uncertainties, such as variations in driving patterns. To achieve this, the entire lifetime of a vehicle is simulated during the optimization. Therefore, a novel methodology is proposed to generate stochastic drive cycles based on application and driver-specific driving profiles. Results show that the generated drive cycles accurately represent driving profiles with respect to the frequency spectra, speed distribution, acceleration distribution, and load characteristics of their corresponding duty cycles. Overall, the novel approach broadens the scope of conventional EV optimization methodologies and therefore increases the significance of results in terms of quality and accuracy. Seizing this opportunity, a variety of studies are performed that reveal significant fuel efficiency gains through driver and application-specific lifetime optimizations.