Computationally Efficient Dynamic Thermal Modeling Based on Dictionary Learning Reconstruction

Accurate and rapid thermal estimation holds immense significance in the analysis of power semiconductors under long-term mission profile, reliability design, and real-time thermal assessment. This letter proposes a novel paradigm shift for thermal estimation of power semiconductors. First, long-term dissipation data are transformed into a limited set of base pulses through orthogonal decomposition. These base pulses are preconverted into corresponding base temperatures, enabling the simplification of long-term thermal estimation by efficient time-shifting and superposition of these base temperatures. Meanwhile, to achieve desired temperature estimation accuracy with a minimal set of base temperatures, we further employ dictionary learning for optimization. To validate the effectiveness of this approach, we compare it against a commercial simulation software and two existing methods. The proposed methodology demonstrates significant advantages in the analysis of long-term mission profile. In addition, we conduct experiments using three distinct standard driving cycles for electric vehicles, all demonstrating the accuracy under highly dynamic loading.