Integrated data analysis approach to select artificial lift method using machine learning.

The optimisation of artificial lift (AL) selection in the oil and gas industry stands as a critical endeavour, directly impacting production efficiency, cost-effectiveness and overall operational success. Traditional AL selection methods rely on engineers' time-consuming field data analysis, which is hindered by data heterogeneity and the complexities of finding meaningful correlations among various parameters, resulting in a universal AL selection gap. This gap has led to AL selection inconstancy, uncertainty in AL parameters screening, production loss due to frequent AL replacement following the installation, and extra expenses. This thesis presents a comprehensive investigation into AL selection, employing innovative machine learning (ML) techniques to upgrade the process by analysing 486,271 data samples, ranging from 2004 to 2021, from 100 wells in a Sudanese oilfield experiencing excessive production loss because of suboptimal AL selection. The study demonstrates the profound impact of ML applications in AL selection, utilizing both supervised learning and clustering techniques. Five supervised ML algorithms are utilised: logistic regression (LR), support vector machines (SVM), K nearest neighbours (KNN), decision tree (DT), and random forest (RF), in addition to K means for clustering. The methodology is applied by developing three distinct ML models, each catering to a unique dataset encompassing production, operation, and environmental/economic parameters. The wells are split into three categories in each model - training, validation and testing - instead of randomly splitting the datasets. This novel methodology streamlines AL selection by expediting data analysis and affording precise results. The outcomes of this research are marked by the remarkable improvements in AL selection accuracy and production performance. Validation of the model using actual field data demonstrated its ability to predict AL and optimal size based solely on production data with over 93% and 92% accuracy. Moreover, the model achieved an accuracy of 91% in predicting the optimal AL using only operational data. Economic and environmental data yielded even higher prediction accuracies, surpassing 99%. Key findings indicate that the predicted MTMPCP and GL outperform the current BPU and NF in terms of production and revenue. Well XFE26 is projected to produce 269 STB/D (equating to over 3 million USD yearly revenue), while Well XJS9 is expected to yield 1878 STB/D (resulting in 11 million USD annual revenue), compared to their current production rates of 97 and 1260 STB/D, respectively. This thesis delves further into identifying the most influential factors affecting AL and size selection. These factors - namely gas, cumulative produced fluid, wellhead pressure, well depth, AL setting depth and AL price - are unravelled through a thorough analysis of the ML models, providing valuable insights into their critical considerations for AL selection in different operational contexts. In conclusion, this thesis serves as a pioneering exploration of ML applications in AL selection, offering tangible solutions to the challenges faced by the industry. The research concludes in a set of robust recommendations. As the oil and gas sector continues to evolve, this research provides a timely and invaluable contribution, pointing the way towards more efficient, cost-effective, and data-driven AL selection practices.

Grantor dc:publisher.institution

Robert Gordon University

Year dc:date.issued

2023

Author dc:creator

• Mahdi, Mohaned Alhaj Abdalla

Advisor dc:contributor.advisor

• M. Amish and G. Oluyemi

• Artificial lift
• Petroleum engineering
• Oil and gas industry
• Machine learning
• Data analysis

Language dc:language

en

Identifier

oai:rgu-repository.worktribe.com:2571156
https://doi.org/10.48526/rgu-wt-2571156

Author Identifier

0000-0002-4153-8734

OAI identifier oai:identifier

oai:rgu-repository.worktribe.com:2571156