Porosity Log Prediction of the Utsira Formation in Sleipner CO2 Storage Site by Implementing Machine Learning Techniques

Abstract

This study focuses on predicting the porosity log of the Utsira sand in the Sleipner CO2 storage site, central North Sea by using a set of machine learning (ML) algorithms. The Utsira formation, a 200-250 m thick late Cenozoic sandstone, has been the main reservoir unit for the Sleipner CO2 storage site. The high porosity (0.35-0.4) of this formation which is overlain by the Nordland shale as the primary seal, makes the Utsira Formation an excellent reservoir to safely store CO2 separated from the produced gas and condensate of the Slepner field. In this study, the porosity was calculated using both the measured density and neutron logs. Then, we incorporated three additional logs such as gamma-ray (GR), sonic (DT), and deep resistivity (RDEP) from a total of nine exploration wells in the area and trained three neural network algorithms of 1) multi-layer feedforward neural network (MLFNN), 2) radial basis function neural network (RBFNN), and 3) deep feedforward neural network (DFNN). Multi-attribute analysis was performed prior to the training phase to ensure better features for the algorithms.

The overall results from the neural networks show better porosity prediction than the multi-attribute analysis. In addition, the DFNN has more robust performance in both training and validation phases compared to the others and, therefore, is the most suitable algorithm to be applied in this dataset. Further work in this study should include more wells in the dataset and attempt to generate more attributes during the training process to enhance the model performance and stability.

Keywords: Reservoir characterization; Porosity log; Sleipner area; Utsira formation; machine learning