Integrating seismic and log data for improved petroleum reservoir properties estimation using non-linear feature-selection based hybrid computational intelligence models

Various petroleum reservoir properties have been estimated in literature using only one of log, seismic or production data. The recent trend in data mining is integrating multi-modal and multi-dimensional data for improved reservoir properties prediction. The objective of this paper is to employ hybrid machine learning and feature-selection based predictive models to estimate the permeability of carbonate reservoirs from integrated seismic and well log data. Hybrid models of Type-2 Fuzzy Logic System (T2FLS) and Support Vector Machine (SVM) with Functional Networks (FN) as the non-linear feature selection algorithm are proposed. Five seismic attributes were integrated with six commonly used Well logs. Data were collected from 33 oil Wells but only 17 of them had a complete matching seismic-log pair. The performance of the hybrid models were compared to those of the individual models without the non-linear but with the conventional feature selection algorithm. The comparative results showed improved prediction accuracies with the hybrid models and more excellently with the FN-SVM model. A blind test on the models revealed that the FN-SVM hybrid model gave an (R-Square) of 0.82, root mean square error of 0.46, and mean absolute error of 0.42 compared to the lowest performing T2FLS model with 0.40, 0.77 and 0.65 respectively. This demonstrates the significance of the hybrid machine learning paradigm in solving petroleum engineering problems with improved accuracies. The study presented some lessons learned from the data limitation challenges experienced in this work and proposed recommendations to chart further research directions.