Determination of Oil-Well Placement at a Petroleum Reservoir using Sequential Convolutional Neural Network Coupled with Robust Optimization under Geological Uncertainty
Abstract
This study proposes a sequential convolutional neural network coupled with robust optimization (SCNN-RO) to determine the optimal placement of an oil production well at a petroleum reservoir under geological uncertainty. The proposed algorithm identifies the well site to maximize the expectation of cumulative oil production for an ensemble (i.e., a set of equiprobable reservoir realizations). The SCNN-RO correlates petrophysical properties near a well as input (e.g., permeability) with cumulative oil production at the well as output. Sequential training is employed for the ensemble to alleviate the computational cost required to acquire training data and determine the optimal well placement using the trained SCNN-RO. The SCNN-RO is operated as follows: First, a single realization is chosen from the ensemble at random and provides training data obtained by running full-physics reservoir simulation for some scenarios installing a well at the realization. Second, SCNN-RO is built by learning training data. Third, the trained proxy is used to estimate cumulative oil production at every candidate well location for all ensemble members and selected prospective well locations based on the ensemble mean of oil productivity estimated using the SCNN-RO. Fourth, full-physics reservoir simulation is performed for the selected scenarios for accurate estimation of oil productivity. We also quantify the similarity between the proxy and reservoir simulation results for assessing the predictability of the proxy. Fifth, the training dataset is updated by adding the simulation results. Lastly, the above processes are iterated until satisfying stopping criteria: threshold for the improvement of proxy performance and maximum iteration number. The performance of SCNN-RO is tested with application to a channelized oil reservoir. We also investigate the efficacies of sequential training and RO on the performance of SCNN-RO. The SCNN-RO can be used as an efficient decision-making tool for solving a well-placement optimization problem balancing with accuracy and computational cost.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMH036.0006K
- Keywords:
-
- 1822 Geomechanics;
- HYDROLOGY;
- 1829 Groundwater hydrology;
- HYDROLOGY;
- 1847 Modeling;
- HYDROLOGY;
- 1849 Numerical approximations and analysis;
- HYDROLOGY