Prediction Of Petroleum Reservoir Properties Using Nonlinear Feature Selection And Ensembles Of Computational Intelligence Techniques

• Main Author: Anifowose, Fatai Adesina
• Format: Thesis
• Language: English
• Published: University Malaysia Sarawak, UNIMAS 2015
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://ir.unimas.my/id/eprint/9327/1/Fatai%20Adesina%20Anifowose%20ft.pdf; http://ir.unimas.my/id/eprint/9327/

Computational Intelligence (CI) techniques have been applied in the prediction of various petroleum reservoir properties but with ample room for improvement. The major objective of the reservoir characterization process is to provide accurate estimates of the reservoir properties to populate full-field simulation models. The recent use of advanced and sophisticated data acquisition tools has led to a data explosion accompanied by very high dimensional data and increased uncertainties. There is the need for robust techniques that will utilize the strengths of some to overcome the weaknesses of others to produce the best results. Despite the persistent quest for better prediction accuracies in the prediction of petroleum reservoir properties, the application of advanced CI methodologies of hybrids and ensembles has either been slowly embraced or not adequately applied. In this thesis, new non-linear feature-selection assisted methods and ensemble learning models are proposed. The algorithms were implemented with optimized tuning parameters and validated with real-life porosity and permeability datasets obtained from diverse and heterogeneous petroleum reservoirs after they have passed on testing them with a benchmark dataset from the UCI Machine Learning Repository. Common metrics were used to evaluate the performance of the proposed models. The standard machine learning paradigm of dividing datasets into training and testing subsets was employed. When implemented on real petroleum engineering datasets, the proposed Functional Networks-Support Vector Machine assisted model attained the highest R-Square performance of 0.96 and 0.87 on the porosity and permeability datasets respectively (compared to the benchmark of 0.90 and 0.82 respectively) and a total execution time of less than 5 seconds. The ensemble models of Artificial Neural Networks with sequentially searched number of hidden neurons, Support Vector Regression with diverse number of the regularization parameter and Extreme Learning Machine assisted with feature selection and randomized assignment of the number of hidden attained the highest R-Square of 0.99 on the porosity and permeability datasets respectively (compared to the benchmark of 0.89 and 0.90 respectively). A thorough analysis of the comparative results showed that our proposed methods and algorithms outperformed the benchmarks. It was concluded that the proposed assisted and ensemble models will significantly increase petroleum exploration efficiency. A number of hybrid and ensemble possibilities have been recommended for future study.