Energy control has been recognized as a core focus in welding research, and the development of a high-accuracy energy prediction model is considered essential for achieving precise energy output in ultrasonic welding. In response to the shortcomings of traditional energy prediction methods, such as difficulties in feature selection, high sensitivity to outliers, and limited prediction accuracy, a prediction model architecture integrating the convolutional neural network (CNN), long short-term memory network (LSTM), and attention mechanism was proposed. A comprehensive review of the physical modeling principles, model design details, and key technologies in ultrasonic energy prediction was presented, covering 1D-CNN feature extraction, bidirectional LSTM temporal modeling, and the design of an attention-enhanced layer for multi-mechanism fusion. The advantages of the CNN-LSTM-Attention model in terms of prediction performance, stability, and generalization ability were analyzed in detail. Experimental results show that this model exhibits outstanding performance in various operating conditions and complex environments, with root mean square error (RMSE) and mean absolute error (MAE) significantly superior to traditional deep learning models.