Use of Machine Learning and Data Analytics to Detect Downhole Abnormalities While Drilling Horizontal Wells, With Real Case Study

Abstract

The standard torque and drag (T&D) modeling programs have been extensively used in the oil and gas industry to predict and monitor the T&D forces. In the majority of cases, there has been variability in the accuracy between the pre-calculated (based on a T&D model) and actual T&D values, because of the dependence of the model’s predictability on guessed inputs (matching parameters) which may not be correctly predicted. Therefore, to have a reliable model, program users must alter the model inputs and mainly the friction coefficient to match the actual T&D. This, however, can conceal downhole conditions such as cutting beds, tight holes, and sticking tendencies. The objective of this study is to develop an intelligent machine to predict the continuous profile of the surface drilling torque to enable the detection of operational problems ahead of time. This paper details the development and evaluation of an intelligent system that could promote safer operation and extend the response time limit to prevent undesired events. Actual field data of Well-1, starting from the time of drilling a 5-7/8-in. horizontal section until 1 day prior to the stuck pipe incident, were used to train and test three models: random forest, artificial neural network, and functional network, with an 80/20 training-to-testing data ratio, to predict the surface drilling torque. The independent variables for the model are the drilling surface parameters, namely: flow rate (Q), hook load (HL), rate of penetration (ROP), rotary speed (RS), standpipe pressure (SPP), and weight-on-bit (WOB). The prediction capability of the models was evaluated in terms of correlation of coefficient (R) and average absolute error percentage (AAPE). The model with the highest R and lowest AAPE was selected to continue with the analysis to detect downhole abnormalities. The best-developed model was used to predict the surface drilling torque on the last day leading up to the incident in Well-1, which represents the normal and healthy trend. Then, the model was coupled with a multivariate metric distance called “Mahalanobis” to be used as a classification tool to measure how close an actual observation is to the predictive normal and healthy trend. Based on a pre-determined threshold, each actual observation was labeled “NORMAL” or “ANOMAL.” Well-2 with a stuck pipe incident was used to assess the capability of the developed system in detecting downhole abnormalities. The results showed that in Well-1, where a stuck pipe incident was reported, a continuous alarm was detected by the developed system 9 h before the drilling crew observed any abnormality, while the alarm was detected 7 h prior to any observation by the crew in Well-2. The developed intelligent system could help the drilling crew to detect downhole abnormalities in real-time, react, and take corrective action to mitigate the problem promptly.