Talara Basin is a complex faulted basin located in the Peruvian northwestern, being developed since the 1890s. However, wells recently drilled in certain areas in the basin were found with the original reservoir pressure. This work presents an application of properly supervised machine learning algorithms (ML) to predict abnormal pressure zones in a study area. For this purpose, input features or parameters were obtained from classified log data commonly used in pore pressure prediction, namely gamma-ray, bulk density, and deep-resistivity. Other estimated parameters such as normal compaction trend values and vertical hydrostatic effective stress have been used for improving the definition of these zones. Four machine learning classification algorithms were trained. Additionally, hyperparameter tunning was performed on each machine learning model to avoid bias and improve accuracy up to 98% on tested data. As well, cross-validation on trained data was performed for enhancing the predictivity of algorithms. Finally, a performance evaluation for the selected machine learning algorithms was introduced based on their ability to predict abnormal pressure zones. Promising results have been obtained using this new approach, showing the technique's applicability to geomechanics for abnormal pressure detection, with the main advantage of using relatively inexpensive wireline log data in problematic areas.
Conventional methods for pore pressure prediction could have limitations, especially in areas with significant structural complexity such as the Talara Basin, located in the Peruvian northwestern. Hence, the use of machine learning algorithms appears as a leading-edge technology for approaching this challenge, not documented before in this area.
Literature works approaching pore pressure estimation with the advantages of machine learning algorithms have been increasing in recent years.
Hu et al., 2013 show the learning powerful capability of a back-propagation neural network for pore pressure with high accuracy.
Naeini et al., 2019 present the use of a supervised deep neural network to predict pore pressure in the Delaware Basin, using compressional and shear velocity, rock density, resistivity, porosity logs, shale volume, and estimated kerogen volume as input data for the prediction, with highly reliable results.