Production Data Analysis by Machine Learning

In this dissertation, I will present my research work on two different topics. The first topic is production data analysis of low-permeability well. The second topic is a quantitative evaluation of key completion controls on shale oil production. In Topic 1, I propose and investigate two novel methodologies that can be applied to improve the results of low-permeability well decline curve analysis. Specifically, I first proposed an iterative two-stage optimization algorithm for decline curve parameter estimation on the basis of two-segment hyperbolic model. This algorithm can be applied to find optimal parameter results from the production history data. By making use of a useful relation that exits between material balance time (MBT) and the original production profile, we propose a three-step diagnostic approach for the preliminary analysis of production history data, which can effectively assist us in identifying fluid flow regimes and increase our confidence in the estimation of decline curve parameters. The second approach is a data-driven method for primary phase production forecasting. Functional principal component analysis (fPCA) is applied to extract key features of production decline patterns on basis of multiple wells with sufficiently long production histories. A predictive model is then built using principal component functions obtained from the training production data set. Finally, we make predictions for the test wells to assess the quality of prediction with reference to true production data. Both methods are validated using field data and the accuracy of production forecasts gives us confidence in the new approaches. In Topic 2, generalized additive model (GAM) is applied to investigate possibly nonlinear associations between production and key completion parameters (e.g., completed lateral length, proppant volume per stage, fluid volume per stage) while accounting for the influence of different geological environments on hydrocarbon production. The geological cofounding effect is treated as a random clustered effect and incorporated in the GAM model by means of a state-of-the-art statistical machine learning method graphic fused LASSO. We provide several key findings on the relation between completion parameters and hydrocarbon production, which provide guidance in the development of efficient completion practices