Temperature estimation of permanent magnet synchronous motor based on neural ordinary differential equation

Permanent magnet synchronous motors (PMSMs) are susceptible to stator winding overheating, which can lead to insulation degradation, short-circuits, or even complete failure. To enhance motor reliability, real-time estimation of stator winding temperature is essential. This study utilizes a Neural Ordinary Differential Equation (Neural ODE) model to analyse the dynamic temperature behavior of PMSM stator windings, incorporating variables such as rotational speed, torque, ambient temperature, and coolant temperature. The model effectively captures the thermal inertia of the PMSM's temperature field, accounting for the cumulative effects of temperature variations. After training, the Neural ODE is employed to predict stator winding temperatures in PMSMs. Performance comparisons with Long Short-Term Memory (LSTM) and Temporal Convolutional Network (TCN) models, under identical data conditions, demonstrate that the Neural ODE model achieves superior prediction accuracy. Specifically, it attains an average mean square error (MSE) of 0.0185, a Mean Absolute Percentage Error (MAPE) of 0.15%, and an R² value of 0.9994 across three experimental settings, highlighting its optimal predictive capabilities.