Convolutional – recurrent neural network proxy for robust optimization and closed-loop reservoir management

Production optimization under geological uncertainty is computationally expensive, as a large number of well control schedules must be evaluated over multiple realizations. In this work, convolutional-recurrent neural network (CNN-RNN) proxy model developed to predict well-by-well oil and water rates, for given time-varying bottom-hole pressure (BHP) schedules, each realization in an ensemble. This capability enables the estimation objective function nonlinear constraint values required robust optimization. The represents extension recently long short-term memory (LSTM) RNN designed rates single geomodel. A CNN introduced here processes permeability realizations, provides initial states RNN. CNN-RNN trained using simulation results 300 different sets BHP We demonstrate accuracy oil-water flow through realizations 3D multi-Gaussian models. then incorporated into closed-loop reservoir management (CLRM) workflow, where it used with particle swarm filter-based method satisfaction. History matching achieved adjoint-gradient-based procedure. shown perform setting five (synthetic) ‘true’ Improved net present value along satisfaction reduction are observed CLRM. For production steps, O(100) runtime speedup simulation-based