Supervised Machine Learning for Liquid Rate Estimation in Oil Field with Limited Well Monitoring Data: A Case Study of the Handil Oil Field

Handil is an oil field located in the Mahakam swamp area. Approximately 80% of its wells are operated with gas lift. Well surveillance primarily relies on pressure, temperature, and gas injection measurements, which are locally updated through daily operator visits. In some wells, an additional single-phase orifice meter is installed on the flow line to estimate gas rates, but it doesn’t provide any information on liquid rates. Mobile testing units, mounted on floating barges, conduct well tests on a campaign basis every 3 to 6 months. However, daily liquid production monitoring becomes extremely challenging due to the lack of metering references at the well level. Furthermore, fast sedimentation area restricts the mobile testing unit to approach the well cluster which further extend the well potential’s uncertainties. Supervised machine learning techniques were employed to estimate the liquid rates of the wells based on available surface parameters. Two machine learning models were developed: liquid rate prediction based on wellhead parameter data (well head estimate) and gas rate correction on water-out gas well based on orifice parameter data (orifice estimate). The preprocessed datasets consisted of clean 930 well test data for the well head estimate and 167 data for the orifice estimate. They were consisted of actual well production test data (wellhead pressure, temperature, gas injection rate, real gas rate, liquid rate, and water-cut), additional flow line parameters for the orifice estimate (flow line pressure, temperature, orifice differential pressure, beta ratio), and the open reservoir depth. The datasets were used to train and test different regression methods, including linear regression, polynomial regression, support vector machines (SVM), and random forest. Random forest regression, providing accuracy up to 93%, was selected on both ML model. Furthermore, additional polynomial regression was applied for model reinforcement, avoiding over-fit and flat prediction on the blind data beyond the training set range. The results show that the developed machine learnings increase the field daily liquid rate production allocation accuracy at 90% from previously only 80% compared to real produced liquid, better detect liquid production decline or incline following dynamic well parameter for quick necessary action, provide new insight to sand production avoidance on specific sand control type, and accurately provide gas correction rate on water-out gas well. This data analytic approach for liquid rate estimation through machine learning contributes to the portfolio of techniques used in the oil and gas industry. It is highly applicable and replicable in oil fields with traditional and limited well monitoring data.