Maximizing the Performance of a Lithium-Ion Battery Aging Estimator Using Reinforcement Learning

As artificial intelligence (AI) technologies have advanced, neural network-based aging estimation has been extensively studied for lithium-ion batteries. Its performance has also been steadily enhanced with sophisticated neural network designs and machine learning skills. However, utilizing well-designed AI-based aging estimators needs more research. This study shows that the same AI-based aging estimator performs differently depending on the input current signal. This article proposes a reinforcement learning (RL)-based framework with digital twin technology using an elaborate electrochemical model strategically to determine an effective input current shape that maximizes the accuracy of a given network-based estimator. As part of RL, the policy network is trained through the digital twin model to generate an input current signal. This enables a more accurate aging estimation, considering the battery's electrochemical states and the initial state of charge (SOC). By specifying a reward for RL training as the estimation error in response to the designed input data, the input current is updated over the training episodes to improve its aging estimation ability. Experimental results show that the RL-based input current signal exhibits an approximately twice more accurate aging estimation than the input signals previously used to train the given neural network-based estimator.