Lost circulation prediction based on machine learning

Abstract

Lost circulation, one of the most headache problems in drilling engineering, causes a significant increase in non-productive time and increases the uncertainty of well control risk. There is a tremendous challenge for lost circulation prediction/diagnosis using traditional methods in carbonate formation due to its complex and variable leakage zone. To solve this problem of the Mishrif reservoir in the H oil field, in this paper, we propose a whole practical workflow of lost circulation prediction based on the Mixture Density Network using deep learning. Firstly, we choose 16 comprehensive mudlogging parameters that are the most correlated with the mud loss rate. These parameters are selected from a total of 22 comprehensive mudlogging parameters using three different feature selection approach. Then, combining the parameters and the mud loss rate, we get the relationship between comprehensive mudlogging parameters and the mud loss rate based on the Mixture Density Network including five sub-Gaussian models. Finally, the mud loss rate distribution is obtained according to the relationship with input parameters in real-time. The reliability is also evaluated using the uncertainty information in the model. The results show that the parameters had the highest correlation with the mud loss rate, including Measure Depth, Vertical Depth, Rate of Penetration, Hook load, Pump Pressure, Stroke Per Minute, Flow In, Flow Line, Temperature In, Temperature Out, Mud weight In, Mud weight Out, Conductivity Out, Equivalent Circulating Density, Total Gas and Pit Volume Total. The Mixture Density Network has a solid ability to describe data suitable for mud loss prediction and diagnosis. Improving the quality of training data is very important to improve the prediction accuracy of the model. Due to uncertainty in the weights of the sub-Gaussian model obtained from training, the final probability density distribution curve may be single or multi-peaked. According to the prediction results, lost circulation can be controlled by the corresponding method. Thus, the workflow not only allows for real-time assessment of lost circulation risk during drilling but also provides a reference for optimizing lost circulation control.