ADAPT seminar

Speaker: Mr. Devesh Kumar (Graduate student, Petroleum and Natural Gas Engineering)
Topic: "Ensemble based assimilation of non-linearly related dynamic data in reservoir models exhibiting non-gaussian characteristics"

Room: 529 Walker Building (refreshments served)
Time: Friday December 1, 2017 3:30pm - 4:30pm

Inverse modeling techniques for estimating reservoir parameters (e.g. Transmissivity, Permeability, etc.) utilize some dynamic (secondary) information (e.g. hydraulic head or production data at well locations) to estimate reservoir parameters. Ensemble based data assimilation methods are one such class of inverse modeling techniques. Ensemble Kalman Filters (EnKF) in specific are built around the basic framework where modeling parameters such as transmissivity, permeability, storativity, porosity, hydraulic head, phase-saturation are included within a state vector psi_f that are updated to psi_a, based on the available dynamic data. Although EnKF presents the ability to update a large number of parameters successively as data becomes available, it suffers from some major drawbacks. It is optimal only in the case when the multivariate joint distribution describing the state vector is Multi-Gaussian. Also the forward model relating the state vector to the observed response is linearized and expressed in terms of the covariance function. These assumptions and simplifications result in models that yield inaccurate predictions of reservoir performance. The aim of this work is to propose a novel method for data assimilation which is free from the Gaussian and linear transfer function assumptions. This new method can be used to sequentially assimilate dynamic data into reservoir models using an ensemble based approach. Updating is performed in the indicator space where modeling is performed non-parametrically and the indicator transform is insensitive to non-linear operations. It is demonstrated that this indicator transform helps us achieve the desired generality which is a shortcoming of EnKF. Because the expected value of indicators directly yields the probability corresponding to an outcome, the method can be used to quantify the residual uncertainty in spatial description of reservoir properties. Because at all steps of the process an ensemble of models is available, so quantification of residual uncertainty in prediction forecasts is possible. Another advantage is that the data assimilation is sequential in nature implying that the updates can be performed in a quasi-real time sense as data becomes available.