Reducing Oil Well Downtime With A Machine Learning Recommender System

Abstract

The oil and gas industry plays an important role in the world’s Gross Domestic Product by providing energy resources to the world. With the price for oil commodities falling in recent years, oil and gas companies require high operational efficiency in order to maintain profits. Unplanned downtime leads to high unnecessary costs representing on average 7.95% of the cost structure of companies in this capitalintensive industry. As a solution, companies have turned to advanced analytics and Big Data to reduce downtime and maintenance costs. This study involves the development of a machine learning recommender system intended to reduce unplanned downtime at oil well facilities. The developed recommender system uses the similarity among customers to predict future purchases and make product recommendations. Predictions are a function of the k-nearest neighbors to each customer, determined using the Euclidean distance or cosine similarity. We followed a binary classification machine learning approach with imbalanced classes by first splitting historical sales data into a training and testing dataset. Then we used the F-2 score and Precision-Recall curve to validate the models’ performance in making accurate recommendations. Recommendations group similar products or services together, reducing the number of times an oil well is taken down for maintenance, therefore reducing downtime. Our results show that this recommender system could lead to a reduction of 1.7 days of downtime and produce cost savings of $2.5 million per customer per year, equivalent to 6.44% savings. The additional products or services sold could lead to additional revenue of $660K per year for the sponsoring company. The recommender system was based on one specific product line within the company, so we believe there is additional opportunity to scale it for larger downtime reduction and increased revenues.