Predicting porosity, permeability and water saturation applying an optimized nearest-neighbour, machine-learning and data-mining network of well-log data

Abstract

Predicting permeability (Ke), water saturation (Sw) and effective porosity (EP) of oil and gas reservoir sections from well log data is necessary task because core data is typically not available for all wells drilled or complete reservoir sections. A novel data matching algorithm is developed to evaluate data from multiple well-log curves. It provides accurate petrophysical metric predictions and extensive data mining insight to the lithological units it evaluates. Expressing the well-log data in a standardized well-log representation and evaluating that network with the data-matching algorithm provides Ke, Sw and EP prediction from datasets that combine standard well logs curves with lithofacies and stratigraphic information. Applying the methodology to a resampled network from a published composite well-log data and lithofacies interpretation (10 variables), for a 100-m section across the Triassic reservoir of the Algerian Hassi R'Mel gas field (Algeria) demonstrates the capabilities of the proposed method. At a screening level of sampling density (~1 sample/10 cm) TOB generates prediction accuracy useful for distinguishing prospective formations (root mean squared errors, RMSE for Ke ~ 15 mD; for Sw ~0.1; for EP ~ 0.01). At a zoomed-interval sampling density (~1 sample/1 cm) for a 10-m zone of interest, TOB provides much improved prediction accuracy (RMSE for Ke ~ 1.3 mD; for Sw ~0.003, for EP ~ 0.0006) for the Hassi R'Mel well-log network. The level of detail provided for each of the two predictions provided by the proposed data-matching method is useful for detailed prediction error analysis. Such data mining capabilities are better than those provided by regression and/or correlation-based machine-learning methods.