A Holistic Data-Driven Framework for Forecasting and Characterization of Tight Reservoirs
Date
2023-12-25
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
In recent years, the need for robust computational techniques in the oil and gas industry, particularly for unconventional reservoirs, has become increasingly evident. This thesis offers a comprehensive exploration into the application of statistical and machine learning (ML), especially deep learning (DL) techniques, with a focus on production forecasting and reservoir classification. The research journey begins by tackling the crucial issue of outlier detection in time-series of production transient data. Analyzing 17 different outlier detection techniques, the study reveals that machine learning-based methods, specifically the k-nearest neighbor and Fulford-Blasingame, outperform other techniques. This not only establishes a reliable data cleaning workflow but also questions the efficacy of traditional statistical methods in dealing with complex, time-dependent datasets. Building upon a clean dataset, the focus then shifts to the limitations of traditional Decline Curve Analysis (DCA) in unconventional tight and shale reservoirs. To address this, the thesis introduces probability density function (PDF)-based models as an alternative to conventional DCA models. These PDF-based models, particularly the Dagum and Beta prime models, have been found to offer superior performance in long-term production forecasting, thus providing a robust toolset for long-term production forecasting and informed decision-making in the industry. In the quest for even more accurate and adaptable solutions, the research leverages deep learning for forecasting long-term well performance. By systematically evaluating a range of recurrent neural networks, the study successfully introduces a versatile model. The proposed Recurrent Neural Networks model for production time series forecasting uses the strengths of RNNs in understanding sequences and trends over time. By using the RNN to understand how these features evolve, the model becomes adept at recognizing complex temporal patterns in time series data, providing a comprehensive framework for accurate forecasting. Finally, the thesis concludes with the introduction of a high-accuracy hybrid deep learning model for reservoir classification. This model, which effectively combines CNNs and Long Short-term Memory (LSTM), has shown an impressive 98% accuracy rate on validation data. It offers a new approach for the automated and accurate classification of unconventional reservoirs, thus setting the stage for its potential as a new tool for reservoir engineers and geoscientist. Throughout this research, several key contributions have been made from establishing a systematic approach for reliable outlier detection to developing high-performance PDF-based and deep learning models for production forecasting and reservoir classification. However, the study also opens avenues for future work, such as the development of labeled real-world datasets, the exploration of PDF-based models across other types of reservoirs, and the incorporation of additional variables and features in deep learning models for even more nuanced analyses. In summary, this thesis serves as a pivotal contribution to both academic research and industrial practice. It not only challenges traditional methodologies but also introduces innovative computational techniques that promise to set a new standard in the field of oil and gas production forecasting and reservoir classification of tight and shale reservoirs.
Description
Keywords
Machine Learning, Decline Curve Analysis, Production Forecasting, Outlier Detection, Reservoir Classification
Citation
Alimohammadi, H. (2023). A holistic data-driven framework for forecasting and characterization of tight reservoirs (Doctoral thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.