Proxy-based well placement optimization

Title: Determination of optimal oil well placement using deep learning under geological uncertainty

• R² of higher than 0.80 was achieved for the first and second six-month periods.
• The proxy could substitute the reservoir simulator in well placement optimization.
• The proxy-based optimization yielded 96% of optimal solution of direct optimization.
• Optimization process was substantially expedited by the proxy-based approach.

Optimal placement of wells is a time-consuming task, which is further made problematic when dealing with geological uncertainties. Although many researchers have attempted to address the intricacies of the problem, they have all taken the single-time optimization of all wells, while the wells are drilled sequentially in real-life fields. This paper presents a viable two-step proxy model to facilitate the process of well placement optimization under geological uncertainty. The proxy model is a deep learning multi-input artificial neural network with convolutional and dense layers, which is used to substitute the reservoir simulator during a one-year field development optimization. Contrary to previous works, this study divides the optimization period into two six-month periods and creates two proxies, where one of them is used for the first period with two wells, and the second proxy is used for the second period but with four wells, to estimate the cumulative oil production as a function of wells location and permeability realizations. After developing and verifying the proxy model, the Particle Swarm Optimization algorithm was coupled with the proxy model to search for the optimal well locations for each period. The proposed approach was tested on two standard benchmark models: The Egg model and the PUNQ-S3 model. Results showed that the proposed proxy model could predict the cumulative oil production with a high coefficient of determination for both the periods and reservoir models on the test subset. More specifically, the coefficient of determination for the Egg and PUNQ-S3 model was 0.89 and 0.89 in the first period, and 0.73 and 0.80 in the second period, respectively. Furthermore, the proxy-based optimization framework could find optimal well locations that yielded a cumulative oil production that was at least 96% of the optimal value obtained from conventional methods for both periods and models, but with approximately 25% and 15% of the time required by the direct optimization approach for the first and second period, respectively. The novelty of this study is to develop separate proxies for sequential well location optimization under permeability distribution uncertainty, which can be used for sequential well placement optimization.

Codes: Simulation and MLP codes

Reservoir models: PUNQ-S3