Selection of Decision Variables for Large-Scale Production Optimization Problems Applied to Brugge Field

Abstract

Abstract Production optimization has got growing interest in the context of closed loop reservoir management during the last few years. A comparative solution project was designed in connection with SPE Applied Technology Workshop in Brugge, June 2008 (Brugge field). The Brugge field data set is used to investigate a large-scale waterflooding optimization problem. The net present value (NPV) is maximized by updating the production and injection rates of the smart completions. We use a set of history matched models generated by Lorentzen et al. (SPE119101) for investigation of the adjoint based waterflooding optimization. Here a large number of decision variables should be updated to maximize the NPV. An efficient procedure is employed to handle the high dimensional optimization problem. The procedure starts with two coarse groups of production/inj ection rates over the time, in the beginning of production and at the moment that the cumulative NPV is halved. After a convergence criterion is met, each coarse group is doubled so that the cumulative NPV values are equal over the refined time interval. The procedure continues until the finest update frequency is achieved. The algorithm is applied to the waterflooding optimization task of the Brugge field. The strategy improves the efficiency of initial optimization steps. Moreover, using the new procedure and 168 variables comparable results were obtained as with optimization using 3,360 variables. Selection of the coarse optimization time intervals for the updating production and injection rates based on the equal cumulative NPV values obtained over the time provides several advantages. First, the strategy decreases the number of the variables tremendously where the large number of optimization variables makes the large-scale optimization algorithms impractical with the current computer resources. Second, larger and faster updates of the rates are made possible as well as achieving higher optimal NPV values for some methods.