Summary
Accurate, data-driven, stochastic models for fluid-flow prediction in hydrocarbon reservoirs are of particular interest to reservoir engineers. Being computationally less costly than conventional physical simulations, such predictive models can serve as rapid-risk-assessment tools. In this research, we seek to probabilistically predict the oil-production rate at locations where limited data are observed using the available data at other spatial points in the oil field. To do so, we use the Bayesian network (BN), which is a modeling framework for capturing dependencies between uncertain variables in a high-dimensional system. The model is applied to a real data set from the Gulf of Mexico (GOM) and it is shown that BN is able to predict the production rate with 86% accuracy. The results are compared with neural-network and co-Kriging methods. Moreover, BN structure enables us to select the most-relevant variables for prediction, and thus we managed to reduce the input dimension from 36 to 17 variables while preserving the same prediction accuracy. Similarly, we use the local-linear-embedding (LLE) method as a feature-extraction tool to nonlinearly reduce the input dimension from 36 to 10 variables with negligible loss in accuracy. Accordingly, we claim that BN is a valuable modeling tool that can be efficiently used for probabilistic prediction and dimension reduction in the oil industry.
Erratum for “Bilevel Data-Driven Modeling Framework for High-Dimensional Structural Optimization under Uncertainty Problems” by Subhrajit Dutta and Amir H. Gandomi
