Neural Operator–Based Thermal Modeling for Electric Motor and Power Electronics Systems

Abstract

Transient finite-element method (FEM) simulations provide detailed temperature fields for electric machines but can be computationally expensive when repeated evaluations are required. This thesis develops a load-memory-enhanced Fourier Neural Operator (FNO) as a computationally efficient surrogate for predicting twodimensional transient temperature fields in a 15 kW squirrel-cage induction motor. The available FEM data are converted from an irregular mesh to a common 512×512 Cartesian grid. The proposed direct model predicts the temperature rise relative to the initial condition using normalized elapsed time, spatial coordinates, a fixed base heat-source field, the instantaneous load, and three exponentially weighted movingaverage load-memory features. These memory features represent short-, medium-, and long-term load exposure without requiring an autoregressive temperature rollout. The model is evaluated through leave-one-condition-out cross-validation on four constant-load conditions and on an independent 100%–50%–75% varying-load trajectory. Across the cross-validation folds, the mean field root-mean-square error is 4.82 ± 2.22 ◦C, with better accuracy for intermediate-load interpolation than for boundary extrapolation. For the varying-load case, the root-mean-square errors of the maximum and spatial mean temperatures are 4.61 ◦C and 3.02 ◦C, respectively. Complete-trajectory inference requires approximately 2.19 seconds, corresponding to an approximately 74-fold measured wall-clock speedup over the reference FEM execution. The results demonstrate the potential of load-memory-enhanced neural operators for rapid transient thermal-field prediction.