The permanent magnet synchronous motor (PMSM) is the core of new energy vehicle drive systems, and its temperature status is directly related to the safety of the entire vehicle. However, the temperature of rotor permanent magnets is difficult to measure directly, and traditional sensor schemes are costly and complex to deploy. With the development of Artificial Intelligence (AI) technology, deep learning (DL) provides a feasible path for sensorless modeling. This paper proposes a prediction model that integrates a Temporal Convolutional Network (TCN), Bidirectional Long Short-Term Memory Network (BiLSTM), and multi-head attention mechanism (MHA) and introduces a Hybrid Grey Wolf Optimizer (H-GWO) for hyperparameter optimization, which is applied to PMSM temperature prediction. A public dataset from Paderborn University is used for training and testing. The test set verification results show that the H-GWO-optimized TCN-BiLSTM-MHA model has a mean absolute error (MAE) of 0.3821 °C, a root mean square error (RMSE) of 0.4857 °C, and an R² of 0.9985. Compared with the CNN-BiLSTM-Attention model, the MAE and RMSE are reduced by approximately 11.8% and 19.3%, respectively.