Single-Cycle Prognostics of Lithium-Ion Batteries Using Statistical Feature Engineering and Extremely Randomized Tree Ensembles

Accurate estimation of State of Health (SoH) and Remaining Useful Life (RUL) is essential for the safe and reliable operation of Lithium-ion Batteries (LiBs). Most data-driven approaches, however, rely on computationally intensive deep learning architectures or require long, continuous degradation histories, limiting their deployment in embedded Battery Management System (BMS). This study presents a lightweight machine learning framework that achieves robust State of Health (SoH) and Remaining Useful Life (RUL) prognostics from statistical features extracted from a single diagnostic charge–discharge cycle. The framework was developed and validated on a publicly available dataset of 124 commercial Lithium Iron Phosphate (LFP)/graphite cells cycled to failure under 72 distinct fast-charging protocols. Sixteen statistical descriptors were computed from the voltage, current, and temperature signals of each cycle. The primary predictive model is a Quantile Extremely Randomized Trees (QERT) ensemble, optimised via Tree-structured Parzen Estimator (TPE) Bayesian search using a generalisation-aware objective. Prediction uncertainty is quantified by applying conformal prediction to the quantile outputs, yielding nominal 95% prediction intervals that achieved empirical coverage of 94.1% for SoH and 92.2% for RUL on the held-out test cells. Permutation-based feature importance analysis reduced the descriptor set to compact subsets of 8 and 10 features for SoH and RUL, respectively. Evaluated on 25 held-out cells unseen during training, the framework achieved aggregate test-set RMSEs of 1.31 percentage points for SoH and 198.87 cycles for RUL, with per-cell median RMSEs of 0.95 percentage points and 89.51 cycles, respectively.