Synergistic enhancement of productivity prediction using machine learning and integrated data from six shale basins of the USA

Abstract

This study aimed to validate the synergistic enhancement of the machine learning model random forest (RF) to predict the oil and gas estimated ultimate recovery (EUR) by integrating well data from basins. The study used data from six shale basins of the USA: Delaware, Marcellus, Barnett, Eagle Ford, Haynesville, and Midland. The input parameters of RF models are composed of fundamental well data such as well location and hydraulic fracturing, which are attainable and feasible to predict EUR. Three cases were proposed to analyze the effect of data integration in this study. Case 1 consisted of independent RF models for each basin; Case 2 involved a single RF model with integrated data of the six basins; and Case 3 used six RF models for the basins based on the RF model of Case 2. The RF models in Case 3 were trained based on the conditions of the learning-completed RF model from Case 2. A comparison of Cases 1 and 2 indicated an improvement in the integration of the well data. Case 2 improved the coefficient of determination in the oil (0.06) and gas (0.03) EUR predictions of the testing data. Additionally, Case 2 reduced the mean square error in the oil EUR of the testing data by 36%. The fine-tuned RF models of Case 3 exhibited better performance than those of Case 2. Based on these results, three key points are highlighted. First, the integrated well data from the six basins succeeds in deriving the synergistic enhancement of RF performance. Second, the combination of data integration and fine-tuning of machine learning models can generally achieve enhanced performance. Third, the proposed EUR prediction based on fundamental well data is expected to be expanded to new well data or even new shale basins, leading to guidelines or productivity maps for the USA shale basins.

Topics & Concepts