A multi-scale data-driven framework for online state of charge estimation of lithium-ion batteries with a novel public drive cycle dataset
Jiaqi Yao, Julia Kowal
Abstract
Lithium-ion batteries are widely used in electric vehicles (EVs) as one of the most promising options with their high energy and power density, where an accurate online state of charge (SOC) estimation is the cornerstone of their safe and optimized usage. This paper proposes a multi-scale data-driven framework for online SOC estimation of lithium-ion batteries, bringing the prior knowledge of battery modeling to data-driven state estimation. The proposed framework utilizes temporal convolutional networks (TCNs) with dilated causal convolutions for multi-scale feature extraction and a cross-scale self-attention feature fusion module proposed for generic 1D feature fusion. A novel public lithium-ion battery drive cycle dataset that covers diverse cycling patterns and ambient conditions is introduced as well, aiming at facilitating the development of SOC estimation models for researchers worldwide. Comprehensive tests and analyses are conducted on two different datasets, demonstrating that the proposed framework outperforms conventional sequence-to-sequence models in most cases. Additionally, the proposed dataset has proven to be highly effective for training and testing SOC estimation algorithms. • A multi-scale data-driven framework is proposed for accurate online SOC estimation of lithium-ion batteries. • A cross-scale self-attention technique is proposed for generic 1D feature fusion. • A novel public lithium-ion battery drive cycle dataset is introduced to facilitate the development of SOC estimation algorithms.