Improved Workflow for Production Modeling in Unconventionals Using Gaussian Mixtures and Principal Components

Abstract

Abstract The current industry-wide practice of generating asset production curves is over-simplified and does not account for a lot of factors. This may lead to reporting errors and challenges in accurately and quickly quantifying well performance and asset potential. The present paper leverages Gaussian Mixture models and principal components to propose a new workflow for production modeling that incorporates all contributory factors while improving accuracy as well as speed. We began by selecting ~2600 gas wells with at least 2 years of production history. Exploratory data analysis was conducted on the geology, petrophysics, well design and completion characteristics of the wells. Gaussian Mixtures were selected as the clustering model due to their performance and synergies with factor distributions. Singular Vector Decomposition was then used to extract the most predictive Eigenvectors (principal components) for each cluster. Cluster-level production profiles are created from these eigenvectors. Thus, this process leverages the predicting factors as well as heterogeneity in each of the well’s production profiles while creating a representative type curve. RMSE values were calculated between the cluster-level predicted production profile and the individual well production curves. GMM-based models performed strongly with an RMSE of 0.146 for the training data and 0.746 for the test data. Additionally, type curves were calculated using more traditional means by taking monthly averages over the region as well as on an operator level. These type curves were then compared to the monthly production values for the populations they represent and the RMSE’s were calculated. The regional type curve had an RMSE of 9.3 and the company-level had an RMSE of 5.9, quantifying the marked improvement from our process. The proposed approach simplifies forecasting by providing rapid, reliable production heuristics for early-life wells without the need for complex, models that may need to be built individually from well to well. The proposed workflow builds upon existing literature on clustering and principal components, to create a novel workflow for reliable and more comprehensive type curve generation. Additionally, it adds to the existing knowledge-based by showcasing how multiple statistical techniques can benefit our modeling work within the industry as well as provide valuable support on early life production forecasting, which is a key challenge.