Abstract
Oil & Gas extraction companies are using a vast amount of capital and expertise on production optimization. The scale and diversity of information required for analysis is massive and often leading to a prioritization between time and precision for the teams involved in the process. This paper provides a success story of how artificial intelligence (AI) is used to dynamically and effeciently optimize and predict production of gas wells. In particular, we focus on the application of unsupervised machine learning to identify under different potential constraints the optimal production parameter settings that can lead to maximum production. A machine learning model is supported by a decision support system that can enhance future drilling operations and also help answer important questions such as why a particular well or group of wells is producing differently than others of the same type or what kind of parameters that work on different wells in different conditions. The model can be advanced to optimize within field constraints such as facility handling capacity, quotas, budget or emmisions.
The methods used were a combination of similarity measures and unsupervised machine learning techniques which were effective in identifying wells and clusters of wells that have similar production and behavioral profiles. The clusters of wells were then used to identify the process path (specific drilling and completion, choke size, chemicals, etc processes) most likely to result in optimal production and to identify the most impactful variables on production rate or cumulative production via an additional clustering of the principle charactersitics of the well. The data sets used to build these models include but are not limited to gas production data (daily volume), drilling data (well logs, fluid summary etc.), completion data (frac, cement bond logs), and pre-production testing data (choke, pressure etc.)
Initial results indicate that this approach is a feasible approach, on target in terms of accuracy with traditional methods and represents a novel, data driven, method of identifying optimal parameter settings for desired production levels; with the ability to perform forecasts and optimization scenarios in run-time.
The approach of using machine learning for production forecasting and production optimization in run-time has immense values in terms of the ability to augment domain expertise and create detailed studies in a fraction of the time that is typically required using traditional approaches. Building on same approach to optimise the field to deliver most reliable or most effeciently against a parameter will be an invaluable feature for overall asset optimisation.
Machine learning for run-time energy optimisation in many-core systems
