Data-Driven Climate Control in Autonomous Electric Vehicles: A Machine Learning Approach to Thermal Comfort and Energy Use

Kurzfassung

Thermal management in battery-electric Level-4 cabins must balance local thermal comfort against HVAC electrical power and the resulting range penalty. This study develops and evaluates a data-driven framework for comfort-oriented and energy-aware steady-state setpoint selection for a hybrid convective–radiant HVAC concept. A vehicle-like mock-up in a climate chamber is instrumented with a 16-zone segmented thermal manikin to measure stationary local equivalent temperatures according to DIN EN ISO 14,505–2 under three ambient scenarios (, , ), two seating postures, and three air-distribution concepts. Based on the resulting steady-state dataset, forward multi-output surrogate models are trained to predict the 16 segmental values from boundary conditions and actuator settings, using a linear regression baseline and ensemble methods (Random Forest, XGBoost). On the pooled dataset with intra-scenario 5-fold cross-validation, XGBoost achieves a mean and a mean across the 16 targets, outperforming the linear baseline (, ). Scenario-level extrapolation remains limited in a leave-one-scenario-out stress test, indicating that explicit climate coverage is required for deployment beyond the represented operating space. The trained surrogate is embedded in a constrained search over the admissible actuation space to identify comfort-feasible setpoints within the Nilsson neutrality bands with minimum electrical HVAC power, computed from a convective energy balance with scenario-dependent COP and measured radiant-panel power. Validation experiments confirm near-neutral comfort in winter and summer. At , radiant support yields only marginal comfort changes but increases electrical HVAC power from about in convection-only operation to about in hybrid operation, causing an estimated WLTP range penalty of roughly in the considered reference cycle. The contribution of the present study is therefore limited to steady-state comfort-constrained setpoint selection and actuator prioritization within the measured operating space and does not extend to broadly generalizable real-time HVAC control.