HMC Techniques for Reducing the Uncertainty of Gas-Lifted Oil Field Model

Parametric model uncertainties could have a high impact on the predictive capabilities of a model. When process measurements become available, these uncertainties may be reduced using parameter estimation techniques. Estimation techniques founded on the Bayesian framework in particular are powerful: they produce a probability density function (PDF) of the estimated parameter rather than a single point estimate.In this paper, we consider a gas lifted oil field model whose predictions are highly sensitive to un-certainty in its parameters. We apply Markov Chain Monte Carlo (MCMC) methods, which follow the Bayesian paradigm, to estimate these parameters, and thereby reduce the uncertainty in the model pre-dictions; two different algorithms, Hamiltonian Monte Carlo (HMC) and No-U-Turn Sampler (NUTS), are used. The probabilistic programming language (PPL), Turing in Julia is used for implementation. Monte Carlo simulations and/or data retrodiction is performed prior to and post parameter estimation, to evaluate the uncertainty in model predictions; the outcomes are compared to determine the efficacy of parameter estimation. Results show that the computed posterior distributions yield model predictions that are in close agreement with the observations, and that model uncertainty is effectively reduced.