A Machine Learning Approach to Predict Gas Oil Ratio Based on Advanced Mud Gas Data

Yang, Tao (Equinor ASA) | Arief, Ibnu Hafidz (Equinor ASA) | Niemann, Martin (Equinor ASA) | Houbiers, Marianne (Equinor ASA) | Meisingset, Knut Kristian (Equinor ASA) | Martins, Andre (Teradata) | Froelich, Laura (Teradata)

OnePetro 

Abstract

Mud gas data from drilling operations provide the very first indication of the presence of hydrocarbons in the reservoir. It has been a dream for decades in the oil industry to predict reservoir gas and oil properties from mud gas data, because it would provide knowledge of the reservoir fluid properties in an early stage, continuously for all reservoir zones, and at low costs. Previous efforts reported in the literature did not lead to a reliable method for quantitative prediction of the reservoir fluid properties from mud gas data. In this paper, we propose a novel approach based on machine learning which enables us to predict gas oil ratio (GOR) from advanced mud gas (AMG) data.

The current work is based on a previous successful pilot in unconventional (shale) reservoirs. Our aim is to extend the results of the pilot study to conventional reservoirs. In general, prediction of reservoir fluid properties is more challenging for conventional reservoirs than for unconventional reservoirs, due to the complexity of petroleum systems in conventional reservoirs. Instead of building a model directly from AMG data, we trained a machine learning model using a well-established reservoir fluid database with more than 2000 PVT samples. After thorough investigation of compositional similarity between PVT samples and AMG data, we applied the model developed from PVT samples to AMG data.

The predicted GORs from AMG data were compared with GOR measurements from corresponding PVT samples to assess the accuracy of the GOR predictions. The results from 22 wells with both AMG data and corresponding PVT samples show large agreement between prediction vs. measurement. The accuracy of the predictive model is much higher than previous results reported in the literature. In addition, a Quality Check (QC) metric was developed to efficiently flag low-quality AMG data. The QC metric is vital to give confidence level for GOR prediction based on AMG data when PVT samples are not available.

The study confirms that AMG data can be used as a new data source to quantitatively predict continuous reservoir fluid properties in the drilling phase. The method can be used to optimize wireline operations and for some cases, it provides a unique opportunity to acquire reservoir fluid data when conventional fluid sampling or use of wireline tools is not possible. After high-quality PVT data becomes available in the wireline logging phase, the continuous GOR prediction can be further improved and used to determine reservoir fluid gradient and reservoir compartmentalization.