Junction temperature is a primary predictor of semiconductor reliability in wind turbine power converters due to its link to thermal expansion driven degradation mechanisms such as solder and bond wire fatigue. Thermal cycling is shaped by environmental inputs (wind speed, ambient conditions) and system operating states (grid loading and control state) across mission profiles. Conventional electrothermal junction temperature estimation requires switching state data or thermal RC networks, or device level measurement, leading to a trade off between speed and accuracy in monitoring. This paper presents a thermal-RC-network-free surrogate workflow in which wind speed is the primary driver of temperature prediction, using surrogate modelling and copulas. A Gaussian copula, as justified by Sklar’s theorem, is trained on moving window wind features and applied using overlapping windows that are recombined into a single time series. The MATLAB implementation further applies anti-overshoot guarding, bias correction, quantile mapping, and temporal alignment. Several post processing stages reference the measured slow temperature level derived from raw junction temperature data for calibration and evaluation. A Clayton copula and monotone mappings provide time-independent dependence modelling between wind and slow junction temperature, together with wind dependent high frequency ripple amplitude modelling elements. The workflow can be used in conjunction with weather forecasting data for short term thermal cycling prediction. The approach is demonstrated on machine and grid side devices of a nominal PMSG 3L NPC system. On the reference profile, the complete MATLAB evaluation path yields RMSE of 0.6928 ◦ C (machine) and 0.1078 ◦ C (grid), with MAPE of 0.35% (machine) and 0.12% (grid).
Bigham et al. (Mon,) studied this question.