A Data-Driven Reinforcement Learning Enabled Battery Fast Charging Optimization Using Real-World Experimental Data
No TL;DR found
Abstract
Fast and safe charging is of great significance for the application of lithium-ion batteries. As one of the most representative optimization decision-making algorithms, reinforcement learning is widely used in fast charging control. However, current research on fast charging based on reinforcement learning uses simulation data for optimization, which deviates from the actual battery. This paper proposes a fast charging strategy to optimize charging time and safety based on reinforcement learning. The agent is trained using the data from real experiment, which includes 600 charge-discharge cycles on each battery. To avoid getting trapped in locally optimal solution, we utilize the $\boldsymbol{\epsilon}$-greedy strategy and discuss the impact of different values of $\boldsymbol{\epsilon}$. Moreover, to avoid dangerous steps in real charging process, we explore the impact of simulating data to real data transfer on training results. We also conduct a comparison experiment to verify the superiority of our strategy. The results show that the strategy we proposed has advantages in terms of charging time, safety, and delaying battery aging. Besides, the transfer of simulation data to real data has been proven to be effective, which can avoid high-risk actions during the exploration process of real experiments and ensure the safety of the experiments.