Early monitoring of gas kick in deepwater drilling based on ensemble learning method: A case study at South China Sea

Gas kick monitoring is of great significance for prevention of blow-out accidents, especially in deep drilling and deep-water drilling. In this study, a machine learning (ML) model for early-monitoring of gas kick is developed using the ensemble learning algorithms based on 7363 lines of drilling logging data at South China Sea. The selected input parameters based on mechanism analysis of gas kick are six fast engineering parameters, including hook load (WHO), weight on bit (WOB), torque (TOR), flow rate (FLW), rate of penetration (ROP) and stand-pipe pressure (SPP), and two slow mud property parameters, i.e. electrical conductivity (CON) and mud outlet density (DEN). The model is constructed using RUSboosted, Subspace-KNN and Bagged Trees algorithms, and is compared with the neural network algorithm. We propose a comprehensive error to quantitatively evaluate the performance of the gas kick monitoring models. The models for early-monitoring of gas kick are applied for a single well and multiple wells, respectively. The results indicate that: (i) The optimal combination of input parameters is made up of six fast engineering parameters and two slow mud parameters. When there is a higher requirement on timeliness, only use of the six fast engineering parameters is also acceptable. (ii) The ensemble learning models work well when the input data expand from single well to multiple wells in the same block. For most cases, the prediction error of the optimal model is below 10%. The RUSboosted algorithm performed best in most data sets. (iii) Gas kick identification from lots of drilling logging records is mathematically a small-sample problem. The output labelling of a potential gas kick should be based on the field practical requirement. The recommended positive length of continuous-point labelling method is 5 m for the studied area, which can effectively reduce the average error from 8.02% to 5.48%.