A stacking ensemble approach for pore pressure prediction in real-time during drilling based on mud log data

The precise and real-time prediction of pore pressure is critical for optimizing drilling efficiency and mitigating the potential risks associated with drilling operations. In order to surmount the limitations of empirical methods and to reduce reliance on logging-while-drilling data, this study pro...

Celý popis

Uloženo v:

Podrobná bibliografie
Vydáno v:	Petroleum science
Hlavní autoři:	Zhang, Dong-Yang, Ma, Tian-Shou, Liu, Yang, Zhang, De-Cheng
Médium:	Journal Article
Jazyk:	angličtina
Vydáno:	Elsevier B.V 01.09.2025
Témata:	Machine learning Mud log Pore pressure Real-time Stacking ensemble approach Stacking ensemble approach Mud log Pore pressure Real-time Machine learning
ISSN:	1995-8226
On-line přístup:	Získat plný text
Tagy:	Přidat tag Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!

Popis
Shrnutí:	The precise and real-time prediction of pore pressure is critical for optimizing drilling efficiency and mitigating the potential risks associated with drilling operations. In order to surmount the limitations of empirical methods and to reduce reliance on logging-while-drilling data, this study proposed a stacking ensemble approach that utilized only conventional mud log data. The workflow involves the preliminary processing of data using isolation forest and wavelet thresholding techniques to effectively eliminate outliers and noise. The Eaton index was estimated using a Bayesian inversion algorithm, and pore pressure was estimated by integrating the dc exponent and Eaton method. A feature selection strategy combining data distribution characteristics and regression-based importance ranking was used to optimize the input parameters. Subsequently, a stacking approach was developed for predicting pore pressure, and the corresponding base learner, meta learner, and hyperparameters were optimized. Finally, the validity of the optimized model was substantiated through field data from three test wells (X1, X5, X3) under different drilling scenarios. The results indicated that the global optimal Eaton index of the block was 0.2449, the maximum percentage error of pore pressure estimation was 3.66%, and the corresponding average error was 1.59% compared to the wireline formation test data. The optimal combination of input features was determined to be R-ROP, WOB, TG, PT, MW, PFI, TVD, H, BR, BT, and SPP. The optimal basic learners were identified as BPNN, CNN, LSTM, and LightGBM, while the optimal meta learners were XGBoost. Prediction accuracy is improved when offset wells are densely distributed, spatially balanced, and proximal to the target well; conversely, sparse or distant offset wells result in reduced prediction performance. The mean absolute percentage errors for test offset well X1, X5, and X3 were 0.4353%, 0.4646%, and 0.6856%, respectively, with the corresponding R2 values of 0.9362, 0.9078, and 0.8950, respectively. Consequently, this approach has the capacity to accurately and in real-time predict pore pressure using solely conventional mud log data. This capability enables timely adjustments to drilling parameters, thereby enhancing operational efficiency and mitigating drilling risks.
ISSN:	1995-8226
DOI:	10.1016/j.petsci.2025.09.020