| dc.description.abstract | Generally, optimum well controls to maximize net present value (NPV) in a waterflooding operation are obtained from a numerical reservoir simulation in combination with an optimization algorithm. This procedure is often computationally expensive and time-consuming to run because of the complexity in numerical reservoir simulation. This complexity also poses a challenge to implement gradient-based optimization algorithms, as there are multiple variables involved and numerous simulations are needed for model evaluations.
This thesis proposes using a machine learning model, specifically an Artificial Neural Network (ANN), to replicate the numerical reservoir simulator outputs. The ANN model is used to predict cumulative oil production, cumulative water injection, and cumulative water production based on sets of well control values, i.e., flowing bottom-hole pressure. Then, the ANN model will be combined with a genetic algorithm (GA) optimization (a derivative-free optimization) to find the optimum well controls that maximize the NPV of a synthetic reservoir model. The optimization results of this model were compared against the results using the open-source FieldOpt software, that optimizes well control values using the genetic algorithm and the reservoir model.
The data generated from reservoir simulation was used as the building blocks to the machine learning model. Hyperparameters optimization was done to create the best machine learning architecture. Several variables were also tested to find the best configuration for the genetic algorithm optimization. The developed ANN model was capable of reproducing the results of the original reservoir model within an acceptable accuracy (1.89%). The genetic algorithm improved the project NPV successfully from the base case and achieved similar results to FieldOpt but 43 hours faster. | |
