A Physically-Constrained Ensemble Learning Rate of Penetration Prediction Model based on Multi-Source Data Fusion

The rate of penetration (ROP) is a key indicator of drilling efficiency in oil and gas extraction. Accurate ROP prediction is crucial for optimizing drilling design and reducing costs. However, current ROP prediction models rarely consider the impact of geological parameters on ROP. Additionally, data-driven models often lack the constraints of physical laws governing the relationships between parameters, leading to poor interpretability of the results. To address these issues, this paper proposes a Correlation-Constrained Ensemble (CCE) model for ROP prediction that integrates both geological and engineering data. The CCE model first uses the Gradient Boosting Decision Tree (GBDT) algorithm to combine four models: Artificial Neural Network (ANN), Regression Tree (RT), Random Forest (RF), and Support Vector Regression (SVR). The model then applies correlation constraints to optimize predictions that do not meet the constraint conditions. By incorporating geological data and applying physical constraints, the mean absolute percentage error (MAPE) of the predictions on the test dataset was reduced from 13.67% to 8.87%. Finally, error evaluation methods more suitable for engineering applications were defined, namely error satisfaction and regional accuracy. Using these methods, the impact of different formations and lithologies on prediction results was analyzed. It was found that the prediction errors are larger at lithological transition interfaces and in formations with high clay content. For the regions with larger prediction errors identified in this study, corresponding solutions for engineering applications have been provided.