Automatic history matching reservoir models using geological features is made challenging by the inability of the modeller to avoid selecting unrealistic reservoir models. There are two main causes for this (1) the selection of unrealistic combinations of geological parameters (e.g. a river channel that is 1 ft wide and 1000 ft deep) and (2) the inability of the modelling method to guarantee the generation of realistic models. In the current practice, of modelling fluvial reservoirs, the geometry of channel sandbodies is based on deterministic or two-dimensional geological priors. These priors can only find relationships between two parameters and lead to a broad field of unrealistic models, which may not be appropriate for a specific case.
We propose a new approach to resolve these problems by: (1) developing robust models of the non-uniform geological parameters space that should be sampled from to find realistic geological models, which could be then incorporated into a framework for automated history matching and uncertainty quantification. (2) Developing a technique that allows us to parameterise a multi-point statistics model of the reservoir based on geological parameters rather than abstract features, like the affinity parameter used in SNESIM.
In this work we show how reservoir models based on realistic geological priors reduce the uncertainty in oil production. We built multi-dimensional geological priors using intelligent techniques, namely Artificial Neural Networks (ANN) and Support Vector Regression (SVR). ANN and SVR allow the priors to capture hidden relations from multiple data sources (modern depositional environments and outcrops). Furthermore, we can predict realistic parameter combinations, not observed in the available data; but still plausible in nature. We sample from the realistic priors within the history-matching framework to achieve the flow responses that match the production data. History-matched models produced under geological realistic conditions reduced the uncertainty in predicting production responses.