Continual learning for online state of charge estimation across diverse lithium-ion batteries
Jiaqi Yao, Bowen Zheng, Julia Kowal
Abstract
An accurate estimation of the state of charge (SOC) ensures the safe and optimized usage of lithium-ion battery systems. With the rapid advances and accelerated iteration of commercial lithium-ion battery cells, the potential benefit for energy storage systems to use different types of battery cells in a mixture emerges. To accommodate this background, this paper proposes a novel continual learning framework for online SOC estimation across diverse lithium-ion batteries. The proposed framework uniquely combines the progressive neural network (PNN) framework with temporal convolutional networks (TCNs), enabling efficient temporal feature extraction while supporting robust forward knowledge transfer and fast adaptation to new battery types without suffering from catastrophic forgetting of the previously learned knowledge. Moreover, in-depth analyses of the interaction between tasks and the influence of task ordering on the model’s performance are provided. Experiment results on the three public lithium-ion battery drive cycle datasets showcase that the proposed continual learning framework is not only able to achieve state-of-the-art accuracy for the SOC estimation of different types of lithium-ion batteries simultaneously, but also has proven to significantly reduce model complexity compared to utilizing multiple conventional single-task models. To the best of our knowledge, this work is the first to apply continual learning to SOC estimation of different types of batteries. Therefore, we believe this work sets up a benchmark for the task of SOC estimation of hybrid lithium-ion battery systems.