Application of Machine Learning in Transient Surveillance in a Deep-Water Oil Field

Abstract

There are on-going efforts in digital transformation in different aspects of hydrocarbon recovery. For well performance surveillance, we have developed the key elements of a Transient Data Surveillance Machine to efficiently process and analyze all transient data from continuous measurements at the wells, allowing for full utilization of the available data. The workflow has been applied at wells in a deep-water oil field in Gulf of Mexico and proved to be effective. We developed Machine Learning (ML) algorithms and techniques to efficiently process and analyze pressure-rate transient data. Following the automatic workflow, K-mean clustering is used to identify shut-in periods, maximum-slope method is used to synchronize pressure and rate data, Supported Vector Machine algorithm combined with Kernel method is used for transient flow-regime recognition, followed by Non-Linear Regression using physical models to estimate reservoir and well properties and assess uncertainty. Through synthetic case and field data testing, we demonstrated that the ML method is tolerant to data noise. Even at 15% of noise level, which is much higher than standard pressure gauge data, the successful rate is 98% in flow-regime identification. However, it is sensitive to data outliers, and we need to include other techniques, such as wavelet data processing, in the workflow. Adding real field data with associated reservoir models that are validated by experts into the training data set could increase the accuracy of pattern recognition 10% more than training with only analytical solutions. The application of our workflow in a deep-water oil field in Gulf of Mexico, which started oil production in 2009 with all wells with permanent downhole pressure gauges, helped to process and analyze transient data from shut-in’s (70% planned transient tests and 30% operation related) efficiently, and derived information about well productivity changes, interference among wells, and permeability reduction due to rock compaction. This enabled continuous well monitoring and effective identification of well productivity issues. The novelty of our Transient Data Surveillance Machine is its capacity in handling huge amounts of dynamic data and its efficiency using real-time data diagnosis for operation decisions and reservoir management.