In this work, we focus on a Bayesian inversion method for the estimation of reservoir properties from seismic data and we study how the inversion parameters, such as rock-physics and geostatistical parameters, can affect the inversion results in terms of reservoir performance quantities (pore volume and connectivity). We apply a Bayesian seismic inversion based on rock-physics prior modeling for the joint estimation of facies, acoustic impedance and porosity. The method is based on a Gibbs algorithm integrated with geostatistical methods that sample spatially correlated subsurface models from the posterior distribution. With the ensemble of multiples scenarios of the subsurface conditioned to the experimental data, we can evaluate two quantities that impact the production of the reservoir: the reservoir connectivity and the connected pore volume. For each set of parameters, the inversion method yields different results. Hence, we perform a sensitivity analysis for the main parameters of the inversion method, in order to understand how the subsurface model may be influenced by erroneous assumptions and parameter settings.
Presentation Date: Monday, October 15, 2018
Start Time: 1:50:00 PM
Location: 206A (Anaheim Convention Center)
Presentation Type: Oral
Developing Oil & Gas assets requires planning production on multiple horizons: (1) the long-term production plan includes strategic decisions for technology and recovery strategies to maximize the Net Present Value (NPV) of the project, (2) the mid-term horizon includes the drilling program and reservoir depletion/injection rates and (3) the short-term optimization (real-time production optimization or RTPO) aims to maximize the usage of the existing facilities. In the case of RTPO, both subsurface and surface systems are important, the goal being to maximize the daily production while honoring all operational constraints.
RTPO requires a comprehensive integrated model covering the entire production system and an accurate mathematical formulation of the problem. This implies finding an appropriate optimization strategy and solver to find an optimal solution within a reasonable time. Sustainable production optimization solutions also assume continuous model update, maintenance and improvement, as the production system behavior changes over time.
In this paper, we develop an integrated model for a complex multi-field asset. The production system includes 12 gas wells, 24 gas-lifted oil wells, 4 gas-injection wells, 4 CO2-injection wells, subsea manifolds, gas pipelines, offshore process facilities and CO2 removal units. Gas production from each field is gathered in a single gas pipeline system connected to a gas processing facility located onshore. Control variables include wellhead pressures, routing of wells, gas lift rates, flaring and re-injection rates. Many capacity, pressure and compositional constraints are considered through the whole production system.
The production optimization model including binary variables and non-smooth non-linear functions is rather challenging to solve. Each part of the integrated model is approximated with multidimensional piecewise-linear functions to a desired degree of accuracy. The resulting Mixed Integer Linear Program (MILP) can be solved efficiently with existing commercial solvers. The optimization solution is used to answer different types of challenge: (1) platform start-up, (2) unexpected failure of a gas compressor, (3) maintenance on a group of wells and (4) changing reservoir conditions. Production increase driven by RTPO ranges from 1 to 5 % with no additional CAPEX. The implementation of the production optimization solution is also discussed. The importance of the usability, user training and solution maintenance is highlighted.
This paper describes how geostatistical inversion based in a Bayesian framework can be modeled and applied on post-stack seismic data, yielding multiple stochastic realizations of acoustic impedance with improved vertical resolution and conditioned to well data. The proposed method is capable to jointly estimate not only the acoustic impedance, but also the wavelet and the uncertainties of the inversion results. The Gaussian assumption for the likelihood models enables to obtain the analytical expressions for the conditioned distributions, which allows sampling from the posterior distribution via Gibbs Algorithm. Here we propose a different convolutional model that simplifies the conditional distributions of the Gibbs algorithm, and discuss in detail how some variables of the stochastic model were defined in a geophysical interpretation. Results of tests on real data are compared with the deterministic Constrained Sparse Spike Inversion and, as expected, clearly show the improvement in the vertical resolution and the conditioning to well data.