Machine Learning Approaches for Pattern Recognition and Missing Data Prediction in Field Datasets from Oil and Gas Operations

Abstract

Abstract The oil and gas industry is currently undergoing a technology transformation with ‘big data’ playing a huge role in making smart data-driven decisions to optimize operations. New tools and systems generate a large amount of data while performing drilling, completions, or production operations and this has become invaluable in well design, field development, monitoring operations as well as optimizing production and recovery. However, sometimes, the data collected has issues that complicate its ability to be interpreted effectively – most commonly being the lack of adequate data to perform meaningful analysis or the presence of missing or null data points. Significant amounts of data are usually generated during the early stages of field development (seismic, well logs, modeling), during drilling and completions (MWD, LWD tools, wireline tools), as well as production operations (production data, pressure, and rate testing). Supervised and unsupervised machine learning (ML) algorithms such as K-Nearest Neighbor, K-Means, Regression (Logistic, Ridge) as well as Clustering algorithms can be used as predictive tools for modeling and interpreting limited datasets. These can be used to identify and resolve deficiencies in datasets including those with missing values and null datapoints. ML and predictive algorithms can be used to determine complex patterns and interdependencies between various variables and parameters in large and complex datasets, which may not be apparent through common regression or curve fitting methods. Work done on a representative dataset of oilwell cement properties including compressive strength, acoustic and density measurements showed potential for accurate pattern recognition with a reasonable margin of error. Missing or null datapoints were rectified through different strategies including interpolation, regression and imputation using KNN models. Supervised machine learning models were determined to be efficient and adequate for structured data when the variables and parameters are known and identified, while unsupervised models and clustering algorithms were more efficient when the data was unstructured and included a sizeable portion of missing or null values. Certain algorithms are more efficient in predicting or imputing missing data values and most models had a prediction accuracy of 85% or better, with reasonable error margins. Clustering algorithms also correctly grouped the datapoints into six clusters corresponding to each class of cement and their curing temperatures, indicating their effectiveness in predicting patterns in unlabeled datasets. Using such machine learning algorithms on oil and gas datasets can help create effective ML models by identifying and grouping similar data with consistent accuracy to complement industry expertise. This can be utilized as a reliable prediction tool when it comes to working with limited datasets or those with missing values, especially when it comes to downhole data.