Abstract
Optimizing well spacing is a critical step in the development of an oilfield. Several approaches have been employed over the past decades in an attempt to obtain an efficient and cost-effective well spacing methodology. The desire to achieve a fast-paced well spacing optimization outcome is not yet realized. This study explores the use of a data-driven surrogate model and metaheuristic optimization algorithms in optimizing the well spacing in a fractured tight reservoir in Saskatchewan, Canada. The wells considered drain oil from the Viking Formation located in the Avon Hill field.
Multilayer perceptron (MLP) neural network is trained to predict time-series cumulative oil and gas production. The training and testing data were obtained from the history matched numerical reservoir model of the Avon Hill field. The artificial neural network (ANN) surrogate model is then coupled with Particle Swarm Optimization (PSO) algorithm, and Genetic Algorithm (GA). Maximization of the net present value (NPV) is set as the objective function with well spacing, fracture conductivity, fracture half-length as the control variables. 3 well spacing optimization workflows are then developed comprising of ANN based proxy only, ANN-PSO and ANN-GA.
It is found that the use of the ANN-based proxy model enables various combinations of well spacing and fracture designs to be tested at higher computational speed compared to numerical simulation. The results also show that the PSO outperformed GA in terms of converging at a higher objective function value and at a higher convergence speed. Applying the proposed workflows to the Avon Hill field, it is found that the optimal well spacing ranges between 124 m to 132 m. This work reveals that variations in fracture geometry design significantly affects hydrocarbon recovery and economic performance of a fractured tight reservoir and these variations can be efficiently explored by the use of machine learning techniques.
