Machine learning for well rate estimation : integrated imputation and stacked ensemble modeling

Abstract

This thesis describes a stacked ensemble, supervised machine learning problem for well rate estimations utilizing well test features that are far from independent and identically distributed (IID), and exhibit missing data with a not missing at random (MNAR) classification from three different oil fields. This research introduces a novel integrated imputation procedure that combines the imputation model selection with the cross-validation procedure for downstream model tuning without data "leakage"--the primary objective shifts from minimizing the imputation data error to minimizing the downstream hold-out error. A stratified time-slicing rolling forecast cross-validation procedure is implemented to minimize over-fitting from the plethora of statistical assumptions that are violated. This thesis seeks to test a framework that will enable well rate estimations for fields available well test data to improve well surveillance capabilities in order to maximize production metrics and minimize adverse health and environmental impacts.