Battery Discharge Capacity Estimation Using Pearson Correlation Retrieval-Based Deep Learning With Initial Cycle Data

초록

In this article, we present a selective-retrieval framework for battery life prediction that, when only a few initial cycles of a new cell are available, trains on a small subset of historically similar cells rather than the full database. Similarity is scored primarily by Pearson correlation (cosine as a fair baseline under identical backbones). On a randomized-policy 18650 set (54 cells) and an external 5 Ah pouch-cell set (12 cells), Pearson-based selection matches or improves accuracy while sharply reducing the amount of training data. On the 54-cell set, Pearson improves both root-mean-square error (RMSE) and mean absolute error (MAE) over training with all data in 46/54 (85.2%) for multilayer perceptron, 52/54 (96.3%) for recurrent neural network, 53/54 (98.1%) for gated recurrent unit (GRU), and 53/54 (98.1%) for Transformer; on the 12-cell validation set, the corresponding fractions are 12/12 (100%), 9/12 (75.0%), 7/12 (58.3%), and 11/12 (91.7%). In the rare instances where Pearson is worse than "All," the increases are small-about 10(-2) Ah in RMSE and 10(-3) Ah in MAE. Using Top-k retrieval also reduces the training set size by similar to 94% (k=3), similar to 91% (k=5), or similar to 87% (k=7) relative to using all 53 source cells, yielding commensurate training time savings. Overall, the method offers a practical, data-efficient update path that preserves (and often enhances) accuracy at substantially lower computational cost.

키워드