Defining 3D Reservoir Well Patterns As Functions Of Geology For Robust Design And Faster Optimization

Abstract

2. Exercise information • Field Development (FD) 3. Optimization Methods : • Latin hypercube sampling on one realization & “selective mean consolidation”. 92 well patterns were created, on a randomly chosen model realization, by Latin Hypercube sampling of the input parameters of a method such that builds patterns obeying a priori optimality rules (see below). A “selective mean consolidation” optimization approach was used to efficiently utilize a preset number of objective functions calls lower than the product of the number of settings multiplied by the number of realizations. 5. Objective function values Field Development (FD): 281 simulations full automatic 1st try: best mean NPV8 610M$, standard deviation 198 M$(while performing qualification tests: best mean NPV8 643M$, standard deviation 245 M$, 481 simulations). 6. Computational complexity to obtain results Field Development (FD) : 281 objective function calls for initial setting exploration and assessing mean performance, ~50% equivalent CPU expenditure in pattern design processes (not yet optimized). Above results reflect the application to the OLYMPUS problem, on first try, of a predefined workflow without any customization to the considered problem (default parameters used for the very few parameters required) and without any optimization process (just sampling). Additional runs were performed to qualify results on some specific aspects. A better result was found (by accident) with NPV8 643 M$ after running 4 separate sets of 50 simulations at a marginal cost in additional non-simulation CPU expenditure. This confirmed, on hindsight, a progress axis relative to the manner in which the problem was formulated from the perspective of drilling costs optimization. Results to date are not deemed likely to be competitive for lack of integration in a recursive global–local optimization process. But they are thought to present an interest as an illustration of the ability of the volunteered problem formulation to reduce the exploration space (thus start from better initial guess and speed up optimization) and reduce the number of objective function calls. Further activity is on-going to demonstrate the potential of the approach when associating it with conventional optimization approaches. Related results might be available for presentation.