A deep learning approach to state of charge estimation of lithium-ion batteries based on dual-stage attention mechanism. (1st April 2022)

• Record Type: Journal Article
• Authors: Yang, Kuo; Tang, Yugui; Zhang, Shujing; Zhang, Zhen
• Is Part Of: Energy. Volume 244(2022)Part B
• Publication Date: 2022-04-01
• Subjects: State of charge -- Dual-stage attention mechanism -- Lithium-ion battery -- Gate recurrent unit -- Public datasets; Power resources -- Periodicals; Power (Mechanics) -- Periodicals; Energy consumption -- Periodicals; 333.7905
• DOI: 10.1016/j.energy.2022.123233
• Languages: English
• ISSNs: 0360-5442

Abstract

Abstract: For lithium-ion batteries, the state of charge (SOC) estimation is one of the most important tasks, and accurate estimation of SOC can provide a guarantee for the continuous operation of electric vehicles. In order to improve the accuracy of SOC estimation and reduce the influence of noise on SOC estimation, a deep learning approach based on dual-stage attention mechanism is proposed. It put features from domain knowledge of lithium-ion batteries such as current, voltage, and temperature, into a gated recurrent unit based encoder-decoder network. In the encoder input stage, we use the input data of the attention mechanism for preprocessing, so that useful features can be adaptively extracted from the input sequence. In the decoder stage, another attention mechanism is used to consider the correlation of the time series, refer to the state of the previous encoder on the time scale, and accurately estimate the SOC at the current moment. The performance of the model is verified on a dataset collected from a lithium-ion battery with various dynamic conditions. The test results show that the proposed method can provide accurate SOC estimation and the mean absolute error can be less than 0.5%. The effectiveness and robustness of the model performance have also been proven on public datasets. Highlights: A deep learning approach based on dual-stage attention mechanism is proposed. A gated recurrent unit encoder-decoder network is used. The model pays more attention to useful information on time and space scales. The model is verified on a dataset collected from various dynamic conditions. The robustness of the model performance has also been proven on public datasets.