Randomized trial examines improved cross-well pore pressure prediction accuracy in oil and gas development, suggesting effective feature selection and model refinements.
Key Points
To enhance the accuracy of pore pressure predictions by addressing feature selection and model architecture for cross-well tasks.
Introduced DASP for feature selection, reducing parameters from 20 to 12
Constructed models including Random Forest, XGB, LGB, and CNN for performance evaluation
Refined CNN architecture by integrating an attention mechanism and physical constraints
LGB showed superior performance during validation, while RF achieved the highest accuracy in cross-well tests
Improved CNN model demonstrated significant enhancements in error metrics and fitting capabilities