Application of GMDH to Predict Pore Pressure from Well Logs Data: A Case Study from Southeast Sichuan Basin, China

• Published in: Natural resources research (New York, N.Y.) Vol. 32; no. 4; pp. 1711 - 1731
• Main Authors: Mgimba, Melckzedeck M., Jiang, Shu, Nyakilla, Edwin E., Mwakipunda, Grant Charles
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.08.2023; Springer Nature B.V
• Subjects: Artificial intelligence; Artificial neural networks; Back propagation; basins; Case studies; Chemistry and Earth Sciences; China; clay; Clay minerals; Computer Science; cost effectiveness; Data structures; Dehydration; Earth and Environmental Science; Earth Sciences; Fossil Fuels (incl. Carbon Capture); Gases; Geography; Group method of data handling; High temperature; industry; Lateral transfers; Learning algorithms; Lithology; Machine learning; Mathematical Modeling and Industrial Mathematics; Mathematical models; Mineral Resources; Neural networks; Original Paper; Parameters; petroleum; Petroleum industry; Physics; Polynomials; Pore pressure; prediction; Pressure; Root-mean-square errors; Statistics for Engineering; Support vector machines; Sustainable Development; temperature; Thermal expansion; Velocity; Well logs; Sichuan Basin; China; Sichuan China; means clustering; GMDH; Well logs; Pore pressure; Machine learning
• DOI: 10.1007/s11053-023-10207-2
• ISSN: 1520-7439

Pore pressure prediction is significant in the petroleum industry because, compared to direct measurement, it is cost-effective and it generates an extensive range of data. Mathematical correlations fail to predict pore pressure due to their failure to include lateral transfer in the reservoir, high temperature and mixed lithology and other mechanisms like aqua-thermal expansion, dehydration of clay and mineral alterations. Also, several machine learning techniques provide unsatisfactory results when predicting pore pressures due to poor selection of input data, over-fitting, slow convergence of results, and manual adjustment of model parameters like hidden layers and weights. To counteract these challenges, we employed, for the first time, group method of data handling (GMDH) technique to predict formation pore pressures from well logs data in the Nanye 1 well, southeast of the Sichuan Basin. Then, the performance of the GMDH technique was compared to other machine learning techniques, including polynomial classifier (POL) and artificial neural networks (ANNs). The GMDH technique provided results with the highest accuracy compared with the other two techniques, giving the lowest root-mean-square error (RMSE) of 0.0308 MPa. In addition, the GMDH technique provided a high coefficient of determination of 0.998. The ANN and POL gave RMSEs 0.0322 and 0.5873 MPa, respectively. Apart from the good results, the GMDH technique was able to identify data structure, direct approximate the results, automatically select the model running parameters and select the relevant input data for predicting the pore pressure, which were the challenges for other techniques. Therefore, the GMDH can be applied to predict pore pressure from the well logs data.