Building fast proxies for optimization of thermal reservoir simulation

With decreasing amounts of conventional energy resources available, production of unconventional oil and gas has become an important new source of energy. Steamassisted gravity drainage (SAGD) is a widely used process, allowing the production of very heavy oils. The numerical simulation of SAGD, however, is complex and highly CPU/time intensive. Moreover, the economical cost of this process is high and production/design optimization could be very beneficial. The high CPU/time requirements, however, render optimization impractical. In this paper, we aim at the creation of fast proxies for SAGD reservoir simulation, with the objective of maximizing net present value (NPV) with respect to selected design parameters. For two practical examples, we use the commercial reservoir simulator STARS to run a number of training simulations. These results are then processed for the creation of an artificial neural network (ANN) and a suport vector regression (SVR) model. Sensitivity analysis is then performed on some of the proxies parameters in order to assess the quality of the models. The key question of how many training runs are required in order to obtain robust proxies for optimization, is also answered with sensitivity analysis. Finally, optimization is performed using the ANN proxy model.