A Novel Error-Bounded Thermal Prediction Methodology in PM Machines via Hybrid CFD and Recurrent Neural Network

Accurate temperature prediction in Permanent Magnet Synchronous Motors (PMSMs) is critical for thermal management and operational reliability, yet existing methodssuch as numerical simulations, lumped-parameter thermal networks (LPTNs), and machine-learning models-face shortages in computational efficiency (numerical simulations), error accumulation (LPTN), and physical interpretability (machinelearning). This study proposes a hybrid methodology integrating computational fluid dynamics (CFD) with thermal neural networks to address disadvantages of various methods in PMSM's temperature prediction. A systematic error analysis reveals divergent dynamics: while recurrent neural networks (e.g., GRU) and LPTN-based models exhibit cumulative errors over time, steady-state CFD simulations provide high-precision, noncumulative results. Leveragicng this insight, the framework periodically injects CFD-derived temperature data into the thermal neural network to reset error-prone hidden states, thereby mitigating long-term error accumulation. Validation using idealized experimental data (substituting CFD results with ground-truth values) demonstrates a 33.7% reduction in rotor temperature prediction RMSE (from 3.53° C to 2.34° C) with 3000- step correction intervals. Results further show monotonic error suppression correlated with correction frequency, highlighting a trade-off between computational feasibility and accuracy. The hybrid approach reduces dependency on training data while enhancing physical interpretability, achieving bounded error propagation under steady-state conditions.