A multi-source domain transfer learning method based on ensemble learning model for lithium-ion batteries SOC estimation in small sample real vehicle data

Hits:

Impact Factor:9.4

DOI number:10.1016/j.energy.2025.137781

Journal:Energy

Key Words:Lithium-ion batteries; Real vehicle data; State of charge; Ensemble learning; Transfer learning

Abstract:State of charge (SOC) is a key parameter in the battery management system, and the accurate estimation of SOC is crucial for the safety and stability of electric vehicles. Aiming at the data scarcity and modelling difficulties caused by small samples and incomplete discharge process under real vehicle conditions, this paper proposes a multi-source domain transfer ensemble learning method based on small-sample real vehicle data. Firstly, a multi-chemical system source domain set covering lithium ternary, lithium iron phosphate laboratory data, and unknown model vehicle data is constructed to enhance the diversity and robustness of transfer. Secondly, the optimal pre-trained neural network models are built for different source domains: BiLSTM model is good at capturing the global temporal pattern, which is suitable for the standard discharge sequence under laboratory conditions; CNN-LSTM combines the local sensing and long-time memory capabilities, which is suitable for the nonlinearity and perturbation in the real vehicle data. Finally, an ensemble learning strategy is proposed to fuse multi-source models and dynamically allocate the weights of each model to effectively avoid the risk of negative transfer. The experimental results show that the method in this paper performs optimally in the SOC estimation task, with an MAE of 0.187 % and an RMSE of 0.245 %, which is significantly better than that of the single-source domain method, and exhibits good generalisation ability and estimation accuracy under multiple battery models.

Note:中科院1区Top

Co-author:Yujing Cai,Chaoxu Mu,Muyao Wu,Heng Li

First Author:Yuan Chen

Indexed by:Journal paper

Correspondence Author:Liping Chen

Document Code:137781

Discipline:Engineering

Document Type:J

Volume:334

ISSN No.:0360-5442

Translation or Not:no

Date of Publication:2025-08-06

Included Journals:SCI、EI

Links to published journals:https://www.sciencedirect.com/science/article/pii/S0360544225034231