Energy-efficient eco-driving for human-driven electric vehicles via adaptive drivability maps with predictive driver modeling

This study proposes a proof-of-concept eco-driving assistance system (EDAS) for human-driven electric vehicles (EVs), based on adaptive drivability maps (ADMs) integrated with predictive driver intent modeling. Differently from existing eco-driving systems that either rely on advisory feedback or apply torque corrections without anticipating human actions, the proposed approach explicitly embeds a preview of the driver's future pedal inputs within a nonlinear model predictive control (NMPC) formulation, where energy efficiency and driver acceptance are jointly addressed at the control level. Driver intent preview is generated through neural network (NN) models, employing two complementary architectures: i) feedforward NNs (FFNNs), offline trained using a physics-based human driver model (PB-HDM) – preliminarily validated on a static driving simulator – for longitudinal control, corresponding to distinct driving styles; and ii) an online-trained long short-term memory (LSTM) NN, continuously updated to capture intermediate or evolving driver behaviors that cannot be strictly categorized. Online switching algorithms select the most appropriate NN configuration in real time. By leveraging driver intent prediction, the proposed ADM framework achieves energy savings through smoother and driver-aware torque corrections. Real-time capable simulation results show that: i) energy consumption is reduced by ∼4% compared to the same vehicle with static drivability maps; and ii) for comparable energy savings, control intrusiveness indicators are approximately halved by the inclusion of driver intent preview, in comparison with the same ADM algorithm excluding the NN-based preview. The results demonstrate the effectiveness of predictive, driver-centered drivability adaptation as a promising and scalable pathway for energy-efficient assistance in human-driven EVs.