Digital twin-supported battery state estimation based on TCN-LSTM neural networks and transfer learning

Zhao, Kai, Liu, Ying ORCID: https://orcid.org/0000-0001-9319-5940, Zhou, Yue ORCID: https://orcid.org/0000-0002-6698-4714, Ming, Wenlong ORCID: https://orcid.org/0000-0003-1780-7292 and Wu, Jianzhong ORCID: https://orcid.org/0000-0001-7928-3602 2025. Digital twin-supported battery state estimation based on TCN-LSTM neural networks and transfer learning. CSEE Journal of Power and Energy Systems 11 (2) , pp. 567-579. 10.17775/CSEEJPES.2024.00900

Preview

PDF - Published Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (2MB) | Preview

Abstract

Estimating battery states such as State of Charge (SOC) and State of Health (SOH) is an essential component in developing energy storage technologies, which require accurate estimation of complex and nonlinear systems. A significant challenge is extracting pertinent spatial and temporal features from original battery data, which is crucial for efficient battery management systems. The emergence of digital twin (DT) technology offers a novel opportunity for performance monitoring and management of lithium-ion batteries, enhancing collaborative capacity among different battery state estimation techniques and enabling optimal operation of battery storage units. In this study, we propose a DT-supported battery state estimation method, in collaboration with the temporal convolutional network (TCN) and the long short-term memory (LSTM), to address the challenge of feature extraction. Firstly, we introduce a 4-layer hierarchical DT to overcome computational and data storage limitations in conventional battery management systems. Secondly, we present an online algorithm, TCN-LSTM for battery state estimation. Compared to conventional methods, TCN-LSTM outperforms other cyclic networks in various sequence modelling tasks and exhibits reduced reliance on the initial state conditions of the battery. Our methodology employs transfer learning to dynamically adjust the neural network parameters based on fresh data, ensuring real-time updating and enhancing the DT's accuracy. Focusing on SOC, SOH and Remaining Useful Life (RUL) estimation, our model demonstrates exceptional results. When testing with 90 cycle data, the average root mean square error (RMSE) values for SOC, SOH, and RUL are 1.1 %, 0.8%, and 0.9 % respectively, significantly outperforming traditional CNN's 2.2%, 2.0% and 3.6% and others. These results un-equivocally demonstrate the contribution of the DT model to battery management, highlighting the outstanding robustness of our proposed method, showcasing consistent...

Item Type:	Article
Date Type:	Publication
Status:	Published
Schools:	Schools > Engineering
Publisher:	Chinese Society for Electrical Engineering
ISSN:	2096-0042
Date of First Compliant Deposit:	3 June 2024
Date of Acceptance:	3 June 2024
Last Modified:	30 Apr 2025 14:45
URI:	https://orca.cardiff.ac.uk/id/eprint/169446

Actions (repository staff only)

Edit Item