A New Proposed Model for Early Kick Detection in Drilling Operation Using Machine Learning

Kick monitoring, detection, and control are key elements to ensure safe drilling operations and avoid catastrophic blow-out incidents that can cause loss of life, equipment, and environmental damage. Conventional kick detection systems such as the pit volume totalizer and the flow in/out sensors identify the kick after a large amount of influx has been recorded on the surface. So, we aim to recognize the kick before it enters the wellbore by detecting the abnormal formation pressure once the bit penetrates the rock. This paper proposes a new machine learning model as an alternative solution using field drilling parameters such as true vertical depth, porosity, bulk density, resistivity, rate of penetration, weight on bit, rotation per minute, torque, standpipe pressure, flow rate, flowline temperature, and gas level. The data-driven models were developed using three separate algorithms: K-Nearest Neighbor, Random Forest, and XG Boost. 6022 field data points were split for training, testing, and validation processes. On average, the model using the random forest algorithm showed the highest accuracy in forecasting the formation pressure, with root mean squared error values and coefficient of determination values of 12.868 and 0.9638, respectively. Streamlit Deployment tool was used to create a user interface program to facilitate the prediction process. The program was tested using 196 field data points and recorded a high accuracy of 95%.

A New Proposed Model for Early Kick Detection in Drilling Operation Using Machine Learning Keywords:

Early Kick Detection , Drilling Parameters , Abnormal Formation Pressure , Machine learning