A novel sorting method for liquid metal batteries based on deep learning and sequential features. (1st August 2023)

• Record Type: Journal Article
• Authors: Xia, Junyi; Shi, Qionglin; Li, Haomiao; Zhou, Min; Jiang, Kai; Wang, Kangli
• Is Part Of: Journal of energy storage. Volume 64(2023)
• Publication Date: 2023-08-01
• Subjects: Battery sorting -- Battery capacity -- Liquid metal battery -- Sequential features -- Deep learning -- One-dimensional convolutional layer; Energy storage -- Periodicals; Energy storage -- Research -- Periodicals; 621.3126
• DOI: 10.1016/j.est.2023.107093
• Languages: English
• ISSNs: 2352-152X

Abstract

Abstract: Energy storage system (ESS) is considered to be an effective solution for renewable energy consumption. Liquid metal battery (LMB), which is a newly emerged battery technology, has great potential in ESS applications and battery sorting is required to improve LMBs' overall performance in the group application. However, current sorting methods that focus on Lithium ion batteries could not meet the requirements of LMBs due to LMB's unique characteristics and data source. Therefore, a rapid battery sorting method based on deep learning and two-dimensional sequential features is proposed in this paper. It adopts a LSTM-CONV1D (long short-term memory unit and one-dimensional convolutional layer) neural network to establish the relationship between the input time series features extracted from the LMB activation dataset and the battery capacity. While the network is trained on segments from discharging curves, the estimation is made based on all the segments of a given discharging curve, which improves the precision of estimation. The model is trained and evaluated on the 50 Ah Li||Sb-Sn LMB's activation dataset, which has achieved an overall RMSE (root mean square error) of 0.3762 on batteries of the test set and an overall 90.77 % accuracy in the sorting application. Since the battery activation dataset utilized in this study is a typical combination of historical data and partial charge-discharge testing data, the proposed sorting method has the potential to be transferred to various forms of datasets such as fast testing (partial constant current charging or discharging test) dataset and different types of batteries. Graphical abstract: Unlabelled Image Highlights: A novel LSTM-CONV1D deep learning neural network is proposed to estimate the sorting index battery capacity. The proposed model directly utilizes sequential features derived from discharging curve. A greedy optimization strategy is adopted for hyperparameters' optimization. A novel segment-based training and cycle-based inference framework is proposed. The proposed model and inference method are validated on batteries of test set and in sorting application.