.png){kind=icon}
.png){kind=icon}
Return
Retrieval-Augmented Degradation Priors for Data-Efficient Early Lithium-Ion Battery Lifetime Prediction
.png){kind=icon}
OA
AI
Y
Y
K
DOI：10.3390/batteries12080284.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Early lithium-ion battery lifetime prediction is difficult because only a small fraction of a cell’s lifetime is observed when a prediction is needed. We evaluate a transparent, validation-tuned combination of a supervised early-cycle predictor and a nearest-neighbor retrieval estimate built from similar historical cells. The method is assessed against supervised-only and retrieval-only baselines over 20 repeated cell-wise splits. On the Severson/Toyota Research Institute/Massachusetts Institute of Technology (Severson/TRI/MIT) dataset (124 lithium iron phosphate/graphite cells), fusion reduces the mean absolute error (MAE) of end-of-life (EOL) prediction by 10.08 cycles at 10 observed cycles and 11.90 cycles at 20 observed cycles relative to the validation-selected supervised baseline; the confidence intervals for these gains exclude zero, whereas gains at 50 and 100 cycles are unsupported. Retrieval-only prediction is competitive in the shortest windows. In a targeted 40-split external confirmation on BatteryLife-XJTU (Xi’an Jiaotong University) at 50 observed cycles, fusion reduces mean MAE from 34.11 to 28.02 cycles (paired bootstrap 95% confidence interval for the gain: 0.47–11.41 cycles; one-sided Wilcoxon p = 0.007 ). The results support retrieval as an interpretable complement to supervised prediction in compatible weak-signal regimes, not as a universal replacement for supervised models.
Keywords:
lithium-ion batteries
battery lifetime prediction
state of health
remaining useful life
retrieval-augmented forecasting
in-context learning
data-efficient learning
time-series forecasting
Journal
B
IF:
4.8
Papers:
1.8K
Citations:
6.9K
Organization
Cited Papers