A Novel Battery Health Prediction Method based on Q-learning Approach

Abstract

Battery health prediction is an important aspect of the battery management system (BMS), which assures safety, reliability, and sustainability in applications such as electric vehicles (EVs). This paper proposes a new technique for battery health prediction using a Q-learning algorithm which is a powerful Reinforcement learning (RL) technique. It is a type of machine learning in which an agent gains decision-making skills through interaction with its environment. The proposed algorithm has been utilized for battery State-of-Health (SoH) estimation in terms of battery capacity which has been considered for computing the State-of-Charge (SoC). It incorporates the optimal tuning of the hyperparameters (i.e. learning rate, discount factor) using Grid search optimization (GSO) within the Q-learning algorithm. The simulation results provide a comparative analysis of the reference SoC computed from the Coulomb counting (CC) method and estimated SoC obtained by the proposed Q-learning algorithm. The proposed approach has been validated on two different NASA battery datasets (B0006 and RW9).