Deriving a Data-Driven Temperature Model for Thermal Operational Safety of Automotive Vehicles

Abstract

Rising vehicle complexity and electrification increase the thermal loads on automotive components, making reliable temperature models essential for ensuring thermal operational safety over the vehicle lifetime. Existing approaches (experimental wind tunnel testing, numerical simulation, and purely data-driven methods) lack scalability to many operating conditions, do not provide physically interpretable parameters, or yield inconsistent results when applied across multiple experiments. This paper addresses the gap of fitting a single, physics-constrained temperature model simultaneously across multiple experimental measurements, enabling consistent parameter estimation and prediction of unseen operating conditions. A lumped parameter thermal network (LPTN) is parameterized using a global minimization approach that classifies each model coefficient as global, discrete-global, or local, depending on whether it is shared across all measurements, across a subset with the same design configuration, or varies individually. The method is evaluated on an electronic control unit (ECU) installed in the BMW 7 Series, using nine wind tunnel measurements covering three different cooling strategies (ventilation, heat pipe, metal insert). A single global model fitted to six measurements achieves a root-mean-square error (RMSE) of 1.09 K, while three unseen measurements are predicted with an RMSE of 1.19 K. Compared to conventional single-measurement fitting, global estimation reduces convergence time to 21.2%, while yielding physically interpretable and consistent parameters across experiments. These results demonstrate that global LPTN parameter estimation provides a fast, robust, and physically interpretable framework for automotive thermal operational safety, capable of reliable extrapolation to unseen conditions with sparse experimental data.