Improving Short-Term and Long-Term Production Forecast Accuracy in Unconventional Reservoirs by Effectively Integrating Static Completion and Well Parameters with Machine Learning and Deep Learning

Accurate production forecasts are vital in the petroleum industry for informed decision-making. However, unconventional reservoirs pose significant challenges due to extended transient flow, operational changes, and limited understanding of the effect of completion and well parameters on production. This dissertation comprises two key phases, “Phase I” and “Phase II”, where Machine Learning (ML) and Deep Learning (DL) models are developed using the Python programming language to improve long-term and short-term production forecast accuracy by incorporating static completion and well parameters into the ML/DL models. “Phase I” develops a practical, data-driven framework using ML/DL models to enhance the accuracy of long-term forecasts in new infill wells producing from the Marcellus formation. Notably, the framework accurately predicts Estimated Ultimate Recovery of infill wells with six-month production history and static completion and well parameters. It also addresses inherent data imbalance issues encountered in oil and gas datasets, which can reduce the performance of the models, and offers several practical solutions to tackle this issue. This framework can be potentially applied to unconventional reservoirs to assist operators throughout the decision-making process to save costs. "Phase II” introduces the Contextual Bi-directional Long Short-Term Memory (C-Bi-LSTM) network to improve short-term production forecast (one month ahead prediction) accuracy during hindcasting in wells producing from Bakken and Three Forks formations. Unlike standard DL models, C-Bi-LSTM can integrate static completion and well parameters known to affect production in unconventional reservoirs into the modeling process. Results demonstrate the superiority of C-Bi-LSTM over standard DL models such as Long Short-Term Memory and Gated Recurrent Unit. Furthermore, “Phase II” models enable simultaneous forecasting of oil, gas, and water rates while also ensuring physically feasible outputs via min-max normalization and sigmoid activation function. C-Bi-LSTM can empower operators with more accurate, stable, and comprehensive production forecasts, enabling them to optimize operations more efficiently and make more-informed production management decisions.

Embargo status: Restricted until 09/2024. To request the author grant access, click on the PDF link to the left.