基于深度嵌入聚类的锂离子电池储能系统不一致性识别

Abstract Inconsistency is an essential cause of weakening the performance of lithium-ion battery packs. Accurate identification of inconsistent batteries is of great significance to the health management of battery energy storage systems (ESSs). Most existing methods require prior knowledge and fail to get optimal representations of dynamic characteristics of batteries, which are no longer suitable for online scenarios with time-varying inconsistency levels. This paper proposes an online unsupervised multi-level inconsistency identification method for battery ESSs based on deep embedded clustering. Firstly, discriminative latent representations are extracted from charge-discharge voltage curves by an improved autoencoder considering both information preservation and reconstruction errors. Secondly, a deep embedded clustering model based on the improved autoencoder and K-means algorithm is built, and then a greedy algorithm is designed to alternately optimize both the latent representations and cluster structures of battery packs without relying on prior knowledge. Thirdly, a distance-based multilevel inconsistency identification framework is constructed for the online consistency management of ESSs. Finally, five months of real-world ESS station data are used to validate the proposed method. The mean clustering inertia indices of our proposed method are respectively 0.9358, 1.1931, 2.1389, and 1.0086 for the four studied battery groups, and the mean Davies-Bouldin indices are respectively 0.7388, 0.7853 0.6396, and 0.6554 for these battery groups, demonstrating higher clustering quality and outperforming other comparative methods. Additionally, compared to the battery management system , the proposed method can identify additional severely inconsistent battery packs within the four battery groups. Furthermore, it has also been successfully applied to a public dataset. All these results prove that the inconsistent batteries can be identified robustly and accurately.