<p>Owing to the unique cooling structure of oil-cooled motors, accurately and rapidly predicting their winding temperature presents a significant challenge in current motor research and development. Conventional deep learning algorithms struggle to satisfy the demand for oil-cooled motor winding temperature prediction under complex working conditions. Therefore, this paper proposes a permanent magnet synchronous motor winding temperature rise prediction model based on a deep learning model combining bidirectional temporal convolutional network, bidirectional long short-term memory network, and attention mechanism. The model utilizes a bidirectional temporal convolutional network to extract hidden information from winding temperature sequences. This extracted information is subsequently fed into a bidirectional long short-term memory network, which has been optimized through an attention mechanism, to perform high-precision predictions. The proposed method achieves high-precision prediction of winding temperature, and the superiority of the proposed model is verified by comparing with other algorithmic models under various working conditions. The results show that the proposed model is better than other models in convergence speed, convergence value, and calculation accuracy, and has good adaptability. Compared with BiLSTM, the convergence speed is increased by 68%, the convergence value is reduced by 61%, MAE is reduced by 39%, RMSE is reduced by 32%, and MAPE is reduced by 44%. Under the WLTC, CLTC, and NEDC operating conditions, the MAE of the predicted temperature of the proposed model is 1.4626, 1.4098, and 1.0038 °C, the RMSE is 2.2427, 1.744, and 1.0038 °C, and the MAPE is 2.18, 2.4831, and 1.62%, respectively.</p>