Multi-scale junction temperature modeling and lifetime assessment of IGBT power modules using graph neural networks

Driven by the global transition toward low-carbon power generation, photovoltaic power generation places increasing reliability demands on power electronic converters. As core components, IGBT power modules experience pronounced multi-scale junction temperature fluctuations under complex operating conditions, leading to thermal fatigue and performance degradation. Accurate junction temperature modeling and reliable lifetime assessment therefore remain critical challenges. This paper proposes a multi-scale modeling framework for IGBT power modules using graph neural networks. The framework integrates module topology and thermal characteristics to jointly model structural, operational, and temporal features. A spatiotemporal graph convolutional network captures inter-chip thermal coupling, while a multi-resolution temporal graph network describes junction temperature evolution across different time scales for lifetime assessment. Experimental results under various grid-connected photovoltaic operating conditions show an average relative error of about 2% in junction temperature prediction, while fatigue damage prediction errors remain within 3%–4%, indicating good accuracy and applicability under complex operating conditions.