Machine Learning Prediction of Regenerative Braking Energy Recovery from Battery Measurement Data

Pack-side current, voltage and battery-state data can quantify charging episodes during driving, but an event-level model must not be credited for reconstructing its target from post-event electrical summaries. We re-examined 8181 negative-current intervals from one Audi e-tron recorded in the San Francisco Bay Area. The original eight-regressor comparison used a partial target-specific feature screen and five-fold gap-grouped and event-ordered validation; those scores are retained only as a shortcut-prone reference. A conservative reanalysis excludes every event-current, mean/end-state and target-derived feature, retaining state of charge, temperature and voltage at the first detected event sample, clock variables, and completed-event duration. Across five 120 s-gap-grouped test folds, the highest mean R2 was 0.511 for energy (Random Forest), 0.132 for peak power (Gradient Boosting), and 0.522 for charge (XGBoost); five forward, gap-group-separated folds provide an additional temporal check. These are post-event estimates rather than forecasts before braking. The 280 gap-defined groups are not verified independent trips, and the dataset contains no brake-pedal, speed, grade or motor-torque signal with which to confirm every charging interval as friction-brake deceleration. No independent vehicle or laboratory experiment was available. The revised benchmark therefore supports only within-vehicle, battery-side inference under the stated feature and grouping rules; it does not support not cross-vehicle deployment or causal claims.