Adaptive sequence-to-sequence learning for long-term permanent magnet synchronous motor temperature prediction

Permanent magnet synchronous motors are widely used in various industrial applications due to their high efficiency and reliable torque generation. However, the permanent magnets in the rotor are highly sensitive to temperature, and continuous thermal stress can cause demagnetization, which leads to performance degradation and a shortened operational lifetime. Therefore, an accurate temperature prediction algorithm is essential for maintaining motor reliability and preventing irreversible magnetic loss. In this work, we propose novel methods for permanent magnet temperature prediction based on a sequence-to-sequence learning architecture. Unlike prior approaches that rely on extensive feature engineering or external environmental variables, the proposed methods employ only control-related electrical inputs together with past temperature values, which improves interpretability and facilitates the design of effective control strategies. Moreover, we incorporate probabilistic teacher forcing and attention mechanisms to further improve the stability and accuracy of long-horizon prediction. Through comprehensive evaluation under realistic operating conditions, we demonstrate that the proposed methods can achieve low prediction error and maintain valid prediction performance over a relatively long forecast horizon.