Abstract: To link historical and predictive data and enhance the accuracy of ultra-short-term top oil temperature predictions for transformers, the predictive performance of various models is compared and a deep learning-based method for ultra-short-term top oil temperature prediction is introduced. The proposed method is founded on the long-short-term memory neural network and Bayesian theory optimization. Data such as the main transformer oil temperature, high-voltage side current, ambient temperature, high-voltage side active power, and relative humidity are utilized for principal component analysis whitening. The deep learning model, optimized for network hyperparameters, follows to capture the ultra-short-term trends, achieving accurate predictions for the transformer’s top oil temperature using deep learning. The Bayesian optimization method efficiently identifies superior hyperparameter combinations with fewer iterations, further minimizing the impact of historical data on the prediction model. The experimental results show that the prediction error of the proposed algorithm model for data within a week is within ±3.5%, and it has better prediction accuracy compared to the other two prediction models.