Abstract: The high rock strength in deep formations leads to low rate of penetration (ROP) and high drilling costs. Accurate ROP prediction is a critical basis for formulating acceleration strategies and optimizing drilling parameters to shorten cycle time and reduce costs while improving efficiency. However, existing data-driven ROP prediction models are mostly trained on fixed historical datasets. During drilling across wells or well sections, data distribution drift occurs due to variations in formations and drilling tools, significantly degrading model generalization capability. Furthermore, existing research on dynamic updating remains fragmented, lacking systematic comparisons of prediction accuracy and update efficiency across multiple models and updating modes. To address these issues, a dynamic updating technical system compatible with data streams while drilling was established for ROP prediction models. First, based on the Young’s ROP equation, raw logging features including weight on bit (WOB), rotary speed, and drilling fluid parameters were screened. Outliers were eliminated using the 3σ criterion combined with an isolation forest, and sliding-window statistical secondary features were constructed to suppress data noise and avoid label leakage. Second, four base models with incremental updating capabilities, namely extreme gradient boosting (XGB), gradient boosting decision tree (GBDT), extreme learning machine (ELM), and deep neural network (knowledge distillation, KD), were selected, and three updating algorithms were proposed: incremental branch growth for tree-based models, recursive least squares parameter iteration for ELM, and transfer learning via knowledge distillation. Five modes of dynamic updating while drilling were designed, and comparative experiments were conducted using 74 711 measured logging data records from 9 wells in the Bohai B Oilfield. The results indicated that incremental model updating in well sections with significant data drift can reduce mean absolute percentage error (E_MAP) by 20%–30%, and inheriting the base model was a necessary condition for ensuring update effectiveness. Simply reusing historical incremental data or performing full retraining tended to introduce data conflicts and reduce prediction accuracy. ELM offered the fastest updating speed but consumed more memory. XGB/GBDT balanced accuracy and computational demands. The knowledge-distilled deep model achieved the highest prediction accuracy but required longer update time. In field applications, the choice of model and updating mode should balance accuracy and timeliness: Incremental updating is recommended for real-time updating while drilling, while offline full-dataset retraining can be employed during idle intervals. ELM is preferred when memory is sufficient, whereas the knowledge-distilled neural network is recommended when high accuracy is pursued. The findings provide theoretical support and engineering references for the adaptive iteration of intelligent ROP prediction models while drilling.