Machine Learning-Based Prediction of Battery Degradation and Remaining Useful Life in Electric Vehicles: A Comparison of Random Forest, Gradient Boosting, and Long Short-Term Memory Models

Abstract

How an electric vehicle (EV) fares in range, safety, resale price, and upkeep comes down to battery degradation, though forecasting that loss remains tough because driving habits and weather conditions interact nonlinearly. Using battery-performance logs of state of charge, charging cycles, voltage, current, temperature, and capacity, this paper builds and benchmarks three machine learning setups—Random Forest, Gradient Boosting, and Long Short-Term Memory (LSTM) networks—to project capacity fade alongside remaining useful life (RUL). We trained the models on engineered cycle-level features, then tested them against held-out cells using mean absolute error (MAE), root mean square error (RMSE), and the coefficient of determination (R²). Operating conditions that drive degradation most heavily were pinpointed through feature-importance analysis, alongside an evaluation of how practical these predictions are for scheduling maintenance. The results, shown as representative figures aligned with patterns seen across the cited literature, show that the LSTM yielded the lowest prediction error for capacity and RUL alike, Gradient Boosting led the non-recurrent options, and Random Forest offered a dependable baseline with lower complexity. Predictive power came primarily from charging-cycle count, operating temperature, and voltage-related metrics, whereas current and state of charge mattered little on their own but raised accuracy when paired with other inputs. These results suggest data-driven RUL forecasts could help steer condition-based maintenance and second-life repurposing, assuming the models undergo testing on field data that captures genuine driving and charging habits.