Reinforcement learning based energy management strategy for fuel cell hybrid vehicles

Official URL: https://risc01.sabanciuniv.edu/record=b2708496_(Table of contents)

There is an increasing concern on the usage of vehicles powered by internal combustion engines due to their high emission levels. The demand for cleaner energy technologies have led to research and development of electric and hybrid vehicles. Among these, fuel cell vehicles have started to draw attention due to the fact that it is clean, sustainable and it has high energy density. Thus, fuel cell hybrid vehicles have the potential to compete with vehicles powered by internal combustion engine in the future, yet there are challenges for fuel cell such as slow dynamics requiring that their operation together should be managed favorably. The primary objective of the thesis is to address the problem of energy management in fuel cell vehicles. For this purpose, first a model of the powertrain is developed. Then, in order to achieve an efficient energy management, a model free reinforcement learning algorithm called deep deterministic policy gradient (DDPG) is employed. The energy management strategy focuses on running the fuel cell in its high efficiency range while limiting the deviation of state of charge of the lithium-ion battery from a target value. It is found that the DDPG agent trained simply with step power inputs can achieve up to 2.7% less energy consumption compared to commonly used rule-based energy management strategies while maintaining the state of the charge of the battery within a certain interval. Our results show that DDPG algorithm shows promising potential to be utilized in such applications.