Ensemble-based big data analytics of lithofacies for automatic development of petroleum reservoirs

Abstract

Big data-driven ensemble learning is explored in this paper for quantitative geological lithofacies modeling, which is an integral and challenging part of petroleum reservoir development and characterization. Quantitative lithofacies modeling involves detection and recognition of underlying subsurface rock’s lithofacies. It requires real-time data acquisition, handling, storage, conditioning, analysis, and interpretation of raw sensory petroleum logging data. The real-time well-logs data collected from the sensor-based tools suffer from complications such as noise, nonlinearity, imbalance, and high-dimensionality which makes the prediction task more challenging. The existing literature on quantitative lithofacies modeling includes several data-driven techniques ranging from conventional well-logs to artificial intelligence (AI). Recently, multiple classifiers based Ensemble learners have been found to be more robust and reliable paradigms for detection and identification tasks in various machine learning applications, however, these are not well embraced in the petroleum industry. Ensemble methodology combines diverse expert’s opinions to obtain overall ensemble decision which in turn reduces the risk of a wrong decision. Thus, the uncertainties associated with complex reservoir data can be better handled by the use of Ensemble learners than the existing single learner based conventional models. Ensemble-based big data analytics, proposed in the paper, includes development and comparative performance testing of five popular ensemble methods (viz. Bagging, AdaBoost, Rotation forest, Random subspace, and DECORATE) for quantitative lithofacies modeling. Seven state-of-the-art base classifiers were used as members of different Ensemble learners for the analysis of Kansas (U.S.A.) oil-field data. The proposed techniques have been implemented on the widely used WEKA platform. The comparative performance analysis of the proposed techniques, presented in the paper, confirms its supremacy over the existing techniques used for quantitative lithofacies modeling.