Machine Learning Applications for Production Prediction and Optimization in Multistage Hydraulically Fractured Wells

Abstract

Due to improvements in horizontal drilling and completion technologies over the past several decades, multistage hydraulic fracturing has become very popular and has led to an explosive growth of the shale and tight oil and gas production worldwide. Even though the completion techniques are well known and relatively simple, the dynamics of fracture formation and hydrocarbon flow within the reservoir are extremely complex. Even with the recent developments, little is known about how the rock mechanical properties, completion design and well spacing affect the morphology of fracture networks and the production of hydrocarbons at the wellhead. Because of this lack in understanding there are no models as of yet that are capable of forecasting production performance with good accuracy. The focus of the thesis is the Montney Formation in Alberta. The research presented in this thesis describes a method to use a convolutional-recurrent neural network (c-RNN) to generate synthetic shear sonic logs with high accuracy and to link a broad range of input parameters, both geological and stimulation, at every stage along a horizontal well bore to the production performance at the well head. The results show that the production performance is driven more by the rock mechanical properties surrounding the perforation clusters than the design of the hydraulic fracture. The results also show that well spacing has affect on production performance. The outcomes of the research provide tools for improving the accuracy of rock mechanical models, optimizing hydraulic fracturing operations with respect to water usage and the placements future wells in the reservoir to maximize gas production.