Production Analysis in Tight/Shale Reservoirs Via Machine Learning Approaches
Date
2024-04-25
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The development of unconventional tight/shale gas reservoirs has undergone a revolutionary transformation, primarily fueled by advancements in horizontal drilling and multi-stage hydraulic fracturing technologies. These innovations have enabled the economical extraction of hydrocarbons from formations characterized by low permeability and complex fracture networks, positioning tight/shale gas as a pivotal component of the global energy mix. The accurate prediction of well production dynamics in these complex formations is a formidable challenge. Traditional empirical and numerical approaches, often fall short due to their inherent simplifications or computational demands. With the surge in data availability from unconventional reservoir developments, leveraging data-driven models for predicting well performance has become increasingly feasible and necessary. This thesis presents a comprehensive suite of machine learning frameworks to predict and enhance well production dynamics in tight/shale gas reservoirs. Initial efforts focus on predicting the first-year cumulative production of infill wells and optimizing their placement and stimulation design. Then, the study delves into the prediction of long-term production dynamics of shale gas wells using a dual-stage attention-based sequence-to-sequence model with some hard physics constraints. By encoding both tabular and time-series data, this model demonstrates superior accuracy and robustness in forecasting well production, outperforming traditional machine learning approaches. Subsequently, a novel physics-informed neural network approach is introduced to deduce the governing partial differential equation for shale gas production decline characterization, integrating Caputo fractional derivatives to capture the heavy-tailed phenomena in production series, thus offering a balance between interpretability and predictive capability. Further, the thesis explores the competitive adsorption of CH4/CO2 in shale formations using a Genetic Algorithm pruned Neural Network. This model robustly predicts the adsorption capacities, offering critical insights for CO2-enhanced shale gas recovery strategies and contributing to carbon capture and storage efforts.
Description
Keywords
Unconventional reservoirs, Machine learning, Production forecast, Physics-informed neural network, Time series
Citation
Wang, H. (2024). Production analysis in tight/shale reservoirs via machine learning approaches (Doctoral thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.