Statistical Causality Analysis: Use Cases In Oil and Gas Applications

Abstract Identifying the underlying causes of issues is important yet difficult. Facilities and operation engineers struggle to determine process variables causing operational difficulties. Causality analysis has made significant advancements in recent years. However, the application in oil and gas field is seldom reported. This study examines the application of statistical and machine learning methods for causality analysis in the oil and gas industry. It includes various use cases to demonstrate the effectiveness of these approaches. This study investigates causality analysis techniques and their applications through comprehensive literature review. The success cases in the oil and gas field are primarily reported using variable importance ranking from tree-based models and word clouds to identify common cases. This study assesses several methods, including the Granger causality test, coefficient ranking from linear models, hierarchical clustering technique, and variable importance ranking from tree-based models. The evaluation incorporates use case evaluation data from various fields and deployment experiences to develop an AI-powered tool that can effectively and efficiently identify the root causes of problems in facilities. This study thoroughly analyses a dataset consisting of shutdown cases and optimization problems. Multiple interviews were conducted with the operations team to identify the actions taken and their outcomes. The underlying causes of each case are pre-determined and stored in the dataset. The data is merged with historical operating data collected from field sensors. Next, statistical and machine learning techniques are applied and conducted back-testing. The method derived from fitting tree-based models or linear models exhibits inherent bias and yields varying outcomes in certain instances. These techniques are primarily applicable to gas processing systems, where the sensor reading responds quickly to changes in process variables. However, they are less effective in liquid-related systems, where the signal response is significantly delayed. However, the Granger causality test has been found to be useful for analyzing liquid cases. Various feature engineering techniques are also examined. Delaying the feature in time is a beneficial technique that enhances the performance of tree-based and linear models. However, it is challenging to determine the appropriate steps for determining the time lag. Multiple methods should be used and considered collectively when analyzing future events. These statistical and machine learning techniques can be used to supplement the root cause analysis activities to promote a more data-driven conclusion in oil and gas businesses. Causality analysis techniques have been found useful for the oil and gas industry in this research. The use cases involve identifying the root causes of unplanned shutdown events and locating the relevant sensors for process optimization. This study's approach, conclusions, best practices, and lessons learned offer substantial advantages for the petroleum industry and academic research, specifically in the fields of data science, artificial intelligence, and facilities and operations.