DEVELOPMENT OF REINFORCEMENT LEARNING BASED CONTROLLER FOR ELECTRIC VEHICLE ENERGY MANAGEMENT WITH HYBRID ENERGY STORAGE SYSTEM

• Main Author: Syaefulah Farsya, Arya
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/52592
• Institution: Institut Teknologi Bandung

The use of combustion engine has become a world serious problem with its emission of dangerous gasses which degrade the air quality and threathen public's health. Amongst the advancing solution is electric vehicle which doesn't directly produce any polutants. However, batteries can degrade over the year which makes ICE vehicles generally last longer than electric vehicle. One of the main reason for such degradation is the high rate or charge/discharge. One solution is to hybridize the energy storage with supercapacitors. Battery has high energy density while having low power density. On the other hand, supercapacitor has high power density but low energy density. This characteristic makes supercapacitor suitable to be used when the vehicle demands high power rate for a short period of time such as when accelerating. With this research, a controller was made to manage the usage of the hybrid energy storage such that the advantages are optimized. Deep deterministic policy gradient, a Reinforcement Learning algorithm was used to train the controller agent. The effect of discount factor and reward function on the agent's performance is also studied. Lastly, the agent's performance is compared to an existing fuzzy logic based controller. It is concluded that the fuzzy logic based controller still outperforms the reinforcement learning based one in terms of maximum battery power rate. It is concluded that the fuzzy logic based controller used as the benchmark still outperforms the reinforcement learning based controller developed in our research in terms of maximum battery power rate minimization. Nevertheless, reinforcement learning method still offer a promising alternative in the design of a decision making agent for its ability to learn strategy without being explicitly programmed.