A series hybrid powertrain provides ultimate freedom in controlling the engine. The flexibility enabled with hybridization creates chances for a synergistic approach, in which the hybrid supervisory control can be augmented to address both the emissions and the efficiency. In this paper, two policy optimization techniques are proposed, namely stochastic dynamic programming and neurodynamic programming, for designing power management controllers. These controllers are then compared with a baseline rule-based controller. The intention is to investigate the additional benefits possible through application of policy optimization algorithms and a systematic framework capable of representing complex system-level effects. The power management of a series hydraulic hybrid vehicle is pursued as a sequential decision-making problem under uncertainty (stochastic control). The low energy density of the hydraulic accumulator adds to the control challenge. First, stochastic dynamic programming and neurodynamic programming are applied to design a controller based on the fuel economy objective. The problem is subsequently expanded to include minimization of transient diesel engine emissions. This poses additional challenges due to the increased state space. The problem is computationally intractable by stochastic dynamic programming and is solved using the newly proposed neurodynamic programming framework. Finally, the supervisory controllers are implemented and evaluated using simulations and an engine-in-the-loop facility. It is shown that, by designing an intelligent multi-objective controller, significant reduction in both the fuel consumption and the emissions can be achieved compared with strategies which focus solely on the fuel consumption.

While much of the previous research relies on Federal Driving Schedules originally developed for emission certification tests of conventional vehicles, consumer acceptance and market penetration will depend on PHEV performance under realistic driving conditions. Therefore, characterizing the actual driving is essential for PHEV design and control studies, and for establishing realistic forecasts pertaining to vehicle energy consumption and charging requirements. To achieve this goal, we analyze naturalistic driving data generated in Field Operational Tests (FOT) of passenger vehicles in Southeast Michigan. The FOT were originally conceived for evaluating driver interaction with advanced safety systems, but the databases are rich with information pertaining to vehicle energy. After the initial statistical analysis of the vehicle speed histories, the naturalistic driving schedules are used as input to the PHEV computer simulation to predict energy usage as a function of trip length. The highest specific energy, i.e. energy per mile, is critical for battery and motor sizing. As an illustration of the impact of actual driving, the low-energy and high-energy driving patterns would require PHEV20 battery sizes of 6.12 kWh and 13.6 kWh, respectively. This is determined assuming that the minimum state of charge (SOC) is 40%. In addition, the naturalistic driving databases are mined for information about vehicle resting time, i.e. time spent at typical locations during the 24-hour period. The locations include ldquohomerdquo, ldquoworkrdquo, ldquolarge-businessrdquo such as a large retail store, and ldquosmall businessrdquo, such as a gas station, and finally ldquoresidentialrdquo other than home. The characterization of vehicle daily missions supports analysis of charging schedules, as it indicates times spent at given locations as well as the likely battery SOC at the time of arrival.

Reinforcement Learning (RL) is gradually being implemented in the Hybrid Electric Vehicle (HEV) supervisory control. Even though RL exhibits significant fuel consumption saving, the long learning time makes it hardly applicable in real-world vehicles. This study aims to reduce the learning iterations of Q-learning in HEV application utilizing warm start methods. Different from previous studies, which initiated Q-learning with zero or random Q values, this study initiates the Q-learning with different supervisory controls, and detailed analysis is given. The results show that the proposed warm start Q-learning requires 68.8% fewer iterations than cold start Q-learning, and improves 10-16% MPG compared to Equivalent Consumption Minimization Strategy control. The results of this study can be used to facilitate the deployment of RL in vehicle applications.