A New Ensemble Machine-Learning Framework for Analyzing EUR of Future Deployed Wells in Tight Gas Reservoir

Abstract

Abstract Tight sandstone gas reservoir is the main type of unconventional natural gas reservoir, and is also the largest unconventional natural gas reservoir in the world. Its significance and role in natural gas resource are becoming more and more obvious[1]. Tight gas reservoirs are generally defined as having less than 0.1 millidarcy (mD) matrix permeability and less than 10% matrix porosity. [2] China's tight gas reservoirs are widely distributed in more than 10 basins such as Ordos, Sichuan, Songliao, Bohai Bay, Qaidam, Tarim and Junggar etc. In recent years, with the advancement and large-scale application of hydraulic fracturing technology, the exploration and development of tight gas reservoir have made significant progress. Two gigantic gas zones in the Sulige Basin in the Ordos Basin and the Xujiahe Formation in the Sichuan Basin was discovered and developed. The prospective resources of low permeability sandstone gas reservoirs exceed 17-24 trillion square meters, accounting for 1/3 of the total natural gas resources in China[3–5]. Nowadays, the main tight gas field of Sinopec has entered a declining stage, facing the challenge of tapping the potential of remaining gas and improving oil recovery. How to analyze the production capacity of future deployed wells and optimize EUR based on clarifying the remaining gas in the gas reservoir is a key issue for efficient development. At present, the evaluation of production capacity and EUR is divided into three categories: theoretical methods (evaluating production capacity based on test data and analyzing EUR based on production data analysis methods, such as RTA)[6–9], empirical methods (such as various experience decline models)[10], and data-driven methods (based on machine learning, neural network algorithms, and other modeling)[11–13]. Although theoretical methods have clear physical meaning, they are difficult to handle the contradiction between the full factor assumptions of physical models and the efficiency of model simulation; Empirical methods establish empirical production decline models based on the analysis of a large amount of production data, but they lack of strict seepage theory support and have poor applicability; Data driven methods mostly reflect the mapping relationship between data and data, and prediction accuracy directly depends on the quality of training data and the suitability of algorithms. The above methods were applicable to gas wells that have been in production for a period of time and show significant production decline characteristic. However, EUR predictions for newly deployed wells in a block are not applicable.