Thermal Neural Networks for Temperature Estimation of Traction Inverters in Electric Vehicles

Abstract:
Efficient energy conversion and enhanced operational durability of competitive traction inverters rely on precise thermal management. As the demand for compact designs in these power-dense devices increases, real-time temperature monitoring becomes critical to mitigate thermal stress. The electrothermal characteristics of traction inverters cover a wide range of time constants — from low thermal mass transistors to high thermal mass capacitors — thus considered a complex inter-domain modeling problem. While first-order principle models, typically represented by high-order partial differential equations, are numerically costly, data-driven approaches render themselves flexible alternatives to achieve accurate yet numerical lightweight models. Hence, this contribution investigates the transferability of the thermal neural network (TNN) method to the temperature estimation problem in traction inverters. The stacked global optimized amplitude time signal (sGOATS) generates a synthetic training dataset, while simulation-based validation is carried out on real-world driving cycles, encompassing both typical operating conditions and challenging edge-case scenarios. Numerical results demonstrate a mean squared error of 2.92 K² across 57 target nodes using a thermally connected dual-model approach with different sampling frequencies, thereby highlighting the scalability and adaptability of TNNs in capturing the diverse thermal dynamics of a traction inverter system.