Machine Learning Outlier Detection Algorithms for Enhancing Production Data Analysis of Shale Gas

Abstract

Economically evaluating shale gas reservoirs, which have huge amounts of reserves, is challenging because of the intricate driving mechanisms. Decline Curve Analysis (DCA) has historically been regarded as the simplest approach for production prediction in shale gas reservoirs since it only requires production history. Nevertheless, uncertainties persist in selecting a suitable DCA model to match the production behavior of shale gas wells. Moreover, the production data are typically noisy due to the dynamic changes in choke size employed to regulate the bottom hole flowing pressure and the periodic shut-ins employed to remove the associated water. Various statistical and machine-learning approaches have been used in the analysis of production forecasting, reservoir property estimation, and resource evaluation. However, many of these methods are not effective in detecting production trends and reservoir signals. The aim of this chapter is to comperhensivly reviewe different machine learning algorithms for outlier detection and assess their efficacy in enhancing the quality of production data for DCA. Out of these algorithms, five were deemed unsuitable since they eliminated entire sections of production data instead of only identifying and eliminating scattered data points. The remaining algorithms (seven) underwent a rigorous evaluation, with a presumption that 20% of the production data is composed of outliers. Further, eight distinct DCA models were studied and implemented before and after removing the noise to test their sensitivity to noise. It was found that usually improving the production data improves their goodness of fitting and reliability of prediction. The clustered-based outlier factor, k-nearest neighbor, and angular-based outlier factor algorithms were found to be effective in improving the data quality for DCA, whereas the stochastic outlier selection and subspace outlier detection algorithms were least effective. Furthermore, certain DCA models, such as Arps, Duong, and Wang models, demonstrated less sensitivity to the removed noise, irrespective of the outlier removal algorithm used. On the other hand, the power law exponential, logistic growth, and stretched exponent production decline models exhibited greater sensitivity to noise removal, with their performance varying based on the employed outlier removal algorithm. The chapter discusses the optimal combination of outlier detection algorithms and DCA models that could potentially mitigate the uncertainties associated with production forecasting and reserve estimation in shale gas reservoirs.