72 real-world BMW trips suggest regenerative braking can spare EV battery range

Keeping an electric car warm draws on the same battery that powers its wheels. Researchers in China have examined how recovered braking energy and machine learning could help manage that demand, using data from 72 real-world BMW i3 trips.

The study was conducted by Baofan Chen of Chongqing Water Resources and Electric Engineering College and Chunrong Zhou of Chongqing Vocational College of Transportation, who compared four algorithms for predicting final battery charge and travel distance.

Their best results came from combining CatBoost with the Horned Lizard Optimization algorithm.

The framework improved prediction performance over the baseline model, while separate modeling examined how regenerative braking energy could support cabin heating.

Building models around real driving

The researchers began with recorded battery and heating circuit data from the BMW i3 journeys.

These measurements captured actual operating conditions, including changing temperatures and driving behavior, rather than relying entirely on standardized laboratory cycles.

They constructed vehicle and heating circuit models, then checked their simulations against the recorded trips to assess whether they reproduced the car’s electrical and thermal behavior.

The team also developed strategies for directing regenerative braking energy into cabin heating. Regenerative braking recovers part of a vehicle’s kinetic energy as electricity during deceleration.

Using that recovered energy for heating can reduce demand on the traction battery. The study reported improvements in driving range and battery life from these strategies, although the supplied account did not specify the size of those gains.

Comparing four prediction algorithms

The machine learning analysis used environmental, vehicle, battery, and thermal system variables to predict final state of charge and distance traveled.

Four models were compared: XGBoost, AdaBoost, CatBoost, and K-Nearest Neighbor. The evaluation combined an 80/20 training and testing split with five-fold cross-validation to assess performance beyond the training data.

CatBoost delivered the strongest baseline results. On the test set, it achieved a coefficient of determination, or R², of 0.7260, alongside a root mean square error of 0.0734 and a mean absolute error of 0.0592. Its cross-validation R² reached 0.802.

The researchers then tested three metaheuristic optimization algorithms to tune CatBoost’s internal settings. Horned Lizard Optimization produced the best-performing combination, named HLO-CatBoost.

Its test R² increased to 0.8408. Root mean square error fell to 0.0559, a reduction of approximately 24 percent, while mean absolute error declined to 0.0434.

Checking reliability across operating conditions

The researchers also examined the framework’s behavior under different environmental scenarios and assessed uncertainty in its predictions.

Computational efficiency analysis evaluated its suitability for practical use. A Williams plot helped identify outliers and establish the model’s applicability domain, indicating where predictions could extend beyond the conditions represented in its training data.

Feature importance analysis examined which inputs most strongly influenced predictions of remaining charge and travel distance.

These checks matter because temperature and cabin heating demand can change how far an EV travels on a charge. More reliable predictions could inform thermal management decisions.

However, the reported accuracy gains describe the model’s predictive performance. They do not quantify an equivalent increase in vehicle range or battery lifespan.

The research was published in the journal Ionics.