Cycle life test optimization for different Li-ion power battery formulations using a hybrid remaining-useful-life prediction method. (15th March 2020)

• Record Type: Journal Article
• Authors: Ma, Jian; Xu, Shu; Shang, Pengchao; ding, Yu; Qin, Weili; Cheng, Yujie; Lu, Chen; Su, Yuzhuan; Chong, Jin; Jin, Haizu; Lin, Yongshou
• Is Part Of: Applied energy. Volume 262(2020)
• Publication Date: 2020-03-15
• Subjects: Li-ion power battery -- Remaining useful life prediction -- Cycle life test optimization -- Deep learning -- Transfer learning -- Arrhenius model; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2020.114490
• Languages: English
• ISSNs: 0306-2619

Abstract

Graphical abstract: Highlights: Test optimization for Li-ion battery formulations reduces the cost of testing. Deep learning is used to predict battery lifespan in high-temperature testing. The transfer learning method is based on a stacked denoising autoencoder. A modified Arrhenius model is used for standard-temperature lifespan estimates. Prediction accuracy is verified and a time savings of nearly 60% is achieved. Abstract: Cycle life testing in battery development is crucial for the selection of a formulation, but it is time-consuming and costly for battery enterprises. A test optimization approach for different Li-ion power battery formulations is designed based on a hybrid remaining-useful-life prediction method to reduce the high cost of constant temperature–stress testing. The test life is replaced by an accurately predicted lifespan to end the testing early and shorten the cycle number. Batteries having the same formulation and tested at different temperatures are integrally optimized for more test savings. Firstly, high-temperature stress testing is stopped early at a preset threshold, and an instance-based transfer learning method is used to predict the battery lifespan by transferring similar historical test samples of different battery formulations to train a highly robust deep learning prediction model. Standard-temperature testing is completely eliminated by utilizing a modified Arrhenius model to estimate the battery lifespan. The model improvements include replacing the high-temperature stress test lifespan with the abovementioned prediction and introducing a prediction error correction coefficient to increase prediction accuracy. The accuracy of the prediction is verified using actual test data from a battery company, resulting in a time savings of nearly 60%. The optimization strategy has extensive application prospects for other constant-stress tests for batteries and other products.