Известия Томского политехнического университета: Инжиниринг георесурсов (Aug 2023)

USING SUPERVISED MACHINE LEARNING ALGORITHMS FOR KICK DETECTION DURING MANAGED PRESSURE DRILLING

  • Roman E. Shcherbakov,
  • Artem V. Kovalev,
  • Andrey V. Ilin

DOI
https://doi.org/10.18799/24131830/2023/8/4125
Journal volume & issue
Vol. 334, no. 8

Abstract

Read online

Link for citation: Shcherbakov R.E., Kovalev A.V., Ilin A.V. Using supervised machine learning algorithms for kick detection during managed pressure drilling. Bulletin of the Tomsk Polytechnic University. Geo Аssets Engineering, 2023, vol. 334, no. 8, рр. 151-163. In Rus. The relevance. The depletion of readily available hydrocarbon reserves determines development of fields with complex geological environment. Managed Pressure Drilling marked the era of high-precision well parameters monitoring during drilling. This technology has provided access to deposits that were previously considered practically «unusable». The main goal of using managed pressure drilling technology is to control downhole pressure within specified limits in order to prevent fluid loss, fracturing, as well as unwanted kick of reservoir fluids into the wellbore. However, if for a certain period of time there is a kick of reservoir fluid from an open borehole or there are losses of drilling fluid, then it is not possible to control the downhole pressure within the specified limits. In this case, it is necessary to use an additional method or algorithm that marks such periods and indicates to the operator or the monitoring system about the presence of kick or absorption of drilling mud. The problems described earlier predetermined the aim of this work. It is claimed that the intelligent system can automatically monitor and analyze parameter trends, detect anomalies in the change of drilling parameters in real time, predict in advance the probability of formation fluid kick and warn the drilling engineer at an early stage, which will allow implementing preventive activity to maintain the required downhole pressure profile. The main aim: create the kick detection machine learning model which predicts kick probability during the managed pressure well drilling using mud logging service data. Objects: multivariate-sensing time-series data of mud logging and measured pressure drilling service. Methods: analysis and evaluation of anomaly detection techniques of determining kick during managed pressure well drilling using machine learning. Results. The authors have performed the overview of anomaly detection techniques of determining kick during managed pressure well drilling using machine learning. Classical machine learning algorithms were tested with labeled test data in order to evaluate its performance. The authors have developed kick detection model with gradient boosting algorithm, evaluated its performance with labeled test dataset. Promising areas of further research were identified.

Keywords