Permanent magnet temperature prediction model based on improved multi-head attention and enhanced loss function

Abstract Accurate prediction of permanent magnet (PM) temperature is crucial for ensuring the safe and stable operation of permanent magnet synchronous motors (PMSMs). However, due to the highly nonlinear coupling among PMSMs variables, extracting spatio-temporal dependencies from multivariate time series is challenging. Additionally, PM temperature variations often involve periodic shocks or pulses, further complicating accurate prediction. To address these problems, this paper proposes a novel PM temperature prediction model based on a long- and short-term time-series improved network (LI), enhanced by a multi-head attention mechanism with modified Softplus nonlinearity bias (MSNB) and an enhanced loss function (EMSE). Key innovations include: (1) MSNB: On the basis of the traditional multi-head attention mechanism (MHA), an improved nonlinear function based on Softplus is introduced as a bias term to model complex nonlinear relationships among PMSMs variables. This enhancement effectively captures spatio-temporal dependencies and automatically focuses on features highly relevant to PM temperature prediction, thereby improving accuracy. (2) EMSE: A novel enhanced mean squared error loss function with an L1 exponential penalty term is designed to dynamically adjust penalties for large errors. Compared to traditional MSE, EMSE strengthens the model’s ability to capture abrupt PM temperature changes, adapts better to complex trends with periodic shocks or pulses, and improves robustness.Experimental results demonstrate that the MSNB-EMSE-LI model significantly reduces the root mean squared error (RMSE) and mean absolute error (MAE) compared to baseline models (MHA and MSE), validating its effectiveness. The proposed model can provide early warnings for potential overheating risks in electric vehicle motors.