Production Forecasting in Tight Gas Reservoirs Using Long Short-Term Memory Methods (LSTM)

Abstract

Forecasting the estimated ultimate recovery (EUR) for extremely tight gas sites with long-term transient behaviors is not an easy task. Because older, more established methods used to predict wells with these characteristics have shown important limitations, researchers have relied on new techniques, like long short-term memory (LSTM) deep learning methods. This study assesses the performance of LSTM estimations, compared to that of a physics-based reservoir simulation process. With the goal of obtaining reliable EUR forecasts, unconventional tight gas reservoir data is generated via simulation and analyzed with LSTM deep learning techniques, tailored for sequential data. Simultaneously, a reservoir simulation model that is based on the same data is generated for comparison purposes. The LSTM forecasting model has the added benefit of considering operational interventions in the well, so that the machine learning (ML) framework is not disrupted by interferences that do not reflect the actual physics of the production mechanism on well behavior. The comparison of the data-driven LSTM deep learning model and the physics-based reservoir simulation model estimations was performed using the latter framework as a benchmark. Findings show that the AI-assisted LSTM model provides predictions similarly accurate to the ones estimated by the physics-based reservoir model, but with the added capability for long-term forecasting. These data-driven EUR models show great promise when analyzing unusually tight gas reservoirs that feature time series well information, which can improve estimations about recovery and point engineers towards better decisions regarding the future of reservoirs. Therefore, exploring deep learning methods featuring varying types of artificial neural networks in greater detail has the potential to significantly benefit the oil and gas sector. When compared to other machine learning methods, novel deep learning techniques have advantages that remain underexplored in the literature. This paper helps fill this gap by providing a valuable comparison between older prediction methods and new estimation simulations based on neural networks that can predict long-term behaviors.