Machine Learning Applications to Improve Pore Pressure Prediction in Hazardous Drilling Environments

Abstract

Pore pressure analysis can be imperative while drilling, especially in certain offshore environments where abnormal pore pressures can cause serious problems such as fluid influx, kicks, and blowouts. In order to avoid such events, the prediction in real-time, or preferably ahead of time, is required. These events can have a catastrophic impact on operations and the safety of the entire platform. Real-time wellbore logs can be used to assist in the prediction of pore pressure, however, due to the high cost of downhole data acquisition and risks associated with a tool getting stuck or lost in hole, comprehensive well logs are not always available. In the absence of a measured sonic log, a predicted acoustic log can be used for input into the pore pressure prediction avoiding the risk and cost of downhole wellbore logs, however, accuracy is extremely important. A Gradient Boosting model is trained and validated on 3 different sets of features to predict acoustic wellbore logs, followed by a physics-based model for pore pressure prediction. The physics-based model is built using Eaton, density extrapolation, and other optimization methods to ensure speed and accuracy. Lastly, the model takes the predicted acoustic data to derive the pore pressure gradient, calibrated by other drilling parameters. The Gradient Boosting model reduces any impact from the lack of data availability, significantly reduces the root mean square error (RMSE) and increases the overall accuracy. The results are then calibrated to drilling events to ensure the predictions are within the range of actual recorded event data. The result from a recent case study shows that pore pressure prediction using predicted acoustic logs correlates closely with recorded drilling events. The client successfully estimated pore pressures using predicted acoustic logs, reducing the for the need to acquire costly logging data downhole. The Gradient Boosting model provides a solution to predict acoustic logs and pore pressure that is highly accurate in real-time. The drilling event calibration method then helps to avoid physical factors that cannot be captured by the model, increasing the overall reliability of the workflow. The method allows for pore pressure analysis to be carried out accurately, regardless of the downhole logs acquired.