Abstract

The objective of this study is to evaluate and compare machine learning techniques for time series models, specifically Long Short-Term Memory (LSTM), with traditional decline curve analysis (DCA) approaches, specifically Arp's Hyperbolic Decline (AHD) method, in forecasting the Estimated Ultimate Recovery (EUR) for two major unconventional reservoirs in the USA, namely Eagle Ford and Bakken. Additionally, this paper aims to discuss the limitations of both methods, such as failing to capture complex temporal patterns and non-linear relationships inherent in unconventional reservoir production data and projection of production multiple years in the future.

The research methodology is designed to be comprehensive and systematic, encompassing data selection, model development, evaluation, and a comparative analysis of results. It utilizes a large dataset comprising production information from 10,000 wells in both the Bakken and Eagle Ford shale formations. The dataset is filtered for wells with a minimum of 36 months of oil production data. The AHD and LSTM models were trained using data from individual wells. These models were then used to forecast production rates over extended periods—30 years for Eagle Ford and 60 years for Bakken—while ensuring a minimum production rate of 4 barrels per day. To measure the precision and efficiency of these predictive models, various evaluation metrics were applied, including the Normalized Root Mean Squared Error (NRMSE) and the Normalized Nash–Sutcliffe Model Efficiency Coefficient (NNSE), providing a quantitative basis for comparison.

In the training phase, LSTM models demonstrated superior ability compared to AHD approach in capturing the complex temporal patterns and nonlinear relationships prevalent in unconventional reservoir production data. This led to significantly improved forecasting accuracy. The notable effectiveness of LSTM models can be attributed to their capacity to learn from extensive datasets and identify intricate patterns that traditional DCA methods often miss. However, despite their strengths, LSTM models showed limitations in projecting long-term declining production rates. They tended to predict rates that plateau, possibly mirroring the last observed rates from the training data, rather than accurately forecasting a decline over several years. Conversely, AHD models managed to maintain a general trend consistent with the data, effectively capturing the decline in production rates over extended periods. For estimating the EUR, the methodology involves summing up all monthly production rates for up to 30 years in Eagle Ford and 60 years in Bakken, or until reaching a minimum oil rate of 4 barrels per day, whichever occurs first.

Both methodologies, however, encounter challenges in accurately projecting production over multiple years especially in the context of operational and completion changes, highlighting the importance of continuous improvement. Ultimately, this comparative analysis enhances our understanding of machine learning and traditional models' capacities and limitations in forecasting EUR for unconventional reservoirs.

This content is only available via PDF.
You can access this article if you purchase or spend a download.