Traditional machine learning methods for mechanical rate of penetration (ROP) prediction are affected by complex feature extraction and imitations of human understanding, which make the prediction accuracy difficult to meet the on-site demand. Based on this, a new ROP prediction method combining feature extraction and regression prediction was proposed. Firstly, data in drilling engineering were pre-processed by box-plot method and one-hot encoding to eliminate abnormal data and to make discrete characteristic continuous. Secondly, convolutional neural network (CNN) was applied to extracting data features, and channel attention mechanism squeeze-and-excitation network (SENet) was introduced to construct a SE-CNN model, which could adjust the importance of CNN feature channel. Finally, SE-CNN model was compared with CNN model. The results show that the goodness of fit of SE-CNN model is increased by 2.1%, the mean absolute error and the root mean squared error are decreased by 1.1% and 1.5%. SE-CNN model has a better prediction accuracy, and can forecast the ROP of drilling field and provides a scientific reference for increasing the ROP during drilling.