Hybrid LSTM–CNN Anti-Jerk Torque Prediction for 2WD and 4WD Electric Vehicles with Single- and Two-Speed Transmissions

Electric powertrains exhibit fast torque dynamics and relatively low mechanical damping, which can excite torsional drivetrain modes and produce longitudinal acceleration oscillations during torque transients. The supplied reference study established a common vehicle-modeling and drivability framework for four central-motor electric-vehicle (EV) architectures: single-speed 2WD (L1), 2-speed 2WD (L2), single-speed 4WD (L3), and 2-speed 4WD (L4). This work extends that framework with data-driven anti-jerk torque prediction using layout-specific hybrid LSTM–CNN models. A hybrid temporal–local feature learning framework is proposed for architecture-aware anti-jerk torque prediction, in which LSTM captures the longer-term temporal evolution of drivetrain states while CNN extracts short-duration local patterns associated with torque transients and torsional oscillations. The proposed controller aims to achieve smoother torque transmission, reduce drivetrain oscillations and vehicle jerk, and improve overall driving comfort. The performance of the system can be evaluated using parameters such as jerk reduction, torque response, and system stability. The proposed hybrid LSTM–CNN algorithm is compared with conventional LSTM to demonstrate its improved anti-jerk torque prediction performance across different EV configurations.