Real-time drilling torque prediction ahead of the bit with just-in-time learning
Kankan Bai, Sheng Mao, Hongbao Zhang, Honghai Fan, Shao-Wei Pan
Abstract
The digital twin, as the decision center of the automated drilling system, incorporates physical or data-driven models to predict the system response (rate of penetration, down-hole circulating pressure, drilling torques, etc.). Real-time drilling torque prediction aids in drilling parameter optimization, drill string stabilization, and comparing the discrepancy between observed signal and theoretical trend to detect down-hole anomalies. Due to their inability to handle huge amounts of time series data, current machine learning techniques are unsuitable for the online prediction of drilling torque. Therefore, a new way, the just-in-time learning (JITL) framework and local machine learning model, are proposed to solve the problem. The steps in this method are: (1) a specific metric is designed to measure the similarity between time series drilling data and scenarios to be predicted ahead of bit; (2) parts of drilling data are selected to train a local model for a specific prediction scenario separately; (3) the local machine learning model is used to predict drilling torque ahead of bit. Both the model data test results and the field data application results certify the advantages of the method over the traditional sliding window methods. Moreover, the proposed method has been proven to be effective in drilling parameter optimization and pipe sticking trend detection. Finally, we offer suggestions for the selection of local machine learning algorithms and real-time prediction with this approach based on the test results.