Abstract: Deep reservoirs in the Bohai Oilfield exhibit high rock strength and significant drilling challenges, making accurate prediction of the mechanical rate of penetration (ROP) critical for drilling efficiency. Variations in well locations and formation depths induce distribution shifts in the input features, limiting the applicability of historical prediction models. To solve the above problem, representative model updating strategies in machine learning and deep learning were reviewed, and a real-time ROP model updating framework was developed based on incremental learning, transfer learning, and knowledge distillation. Comparative evaluations of baseline models, including Extreme Learning Machine (ELM), Gradient Boosting Tree, and Deep Neural Network, under different updating schemes indicate that the improvement brought by model updating becomes more pronounced as data drift increases. Compared with full retraining using all available data, incremental updating achieves superior overall performance by maintaining a favorable balance between prediction accuracy and computational efficiency. Although each baseline model exhibits distinct strengths and limitations, ELM demonstrates the best overall performance, reducing the mean absolute percentage error by 20%–30% in wells where model updating is effective. These results indicate that dynamic model updating can effectively mitigate performance degradation caused by data drift. The proposed approach provides robust computational support for drilling optimization and promotes the development of intelligent drilling technologies.