Comparative Study of Novel Population-Based Optimization Algorithms for History Matching and Uncertainty Quantification: PUNQ-S3 Revisited

Abstract

Abstract In modern field management practices, there are two important steps that shed light on a multimillion dollar investment. The first step is history matching where simulation model is conditioned to production and/or seismic data. In this inverse problem, we calibrate our model to reproduce the historical observations from the field. In second step we quantify uncertainty of the predictions made by calibrated models. These two steps are tied together; multiple history matched models are essential for a realistic uncertainty estimate of the future field behavior. Stochastic population-based optimization methods have been used in the last two decades as popular tools in history matching frameworks. Stochastic sampling algorithms explore/exploit the parameter space to find diverse good-fitting models. Recently two innovative algorithms were proposed to tackle history matching problem; ant colony optimization [Hajizadeh et al. 2009] and differential evolution [Hajizadeh et al. 2010]. However these algorithms were applied for history matching of a simple reservoir model with few unknown parameters. The question is the capability of these new methods for solving complex history matching problems and estimation of the uncertainty associated with these models. This paper compares the application of ant colony, differential evolution and neighbourhood algorithms for history matching and uncertainty quantification of the PUNQ-S3 reservoir. PUNQ-S3 model is a synthetic benchmark case with challenging parameterization, history matching and uncertainty quantification steps. We compare performance of the above algorithms in sampling the parameter space and obtaining multiple history-matched models. The paper also includes comparison of convergence properties of these algorithms for this high dimensional problem. We show that novel stochastic population-based optimization algorithms can be successfully applied for history matching problems with large number of unknown parameters. The algorithms are integrated with a Bayesian framework to quantify uncertainty of the predictions. Results confirm that the proposed methodology provides reliable predictions of the future reservoir recovery.