Predicting System Surface Parameters Using Artificial Neural Network

Abstract

Abstract The prediction of apparent surface torque and the system standpipe pressure holds immense importance in any automated system or digital twin solution. These parameters provide crucial insights that are instrumental in determining various factors in the digitalized drilling application workspace. However, obtaining these values prior to the operation poses a challenge due to their dependence on numerous lithological and operational parameters. Due to the problem of non-linearity, a statistical tool is favored in developing a prediction system for these features. Artificial neural networks (ANN), a statistical tool in machine learning (ML), can effectively predict the system standpipe pressure and the apparent surface torque. A logical base data cleaning process is conducted to ensure consciousness cleaning of the dataset based on statistical feature exploration, feature engineering, and domain knowledge. A large dataset of 336 wells from a single operator across four concessions is used to train the ANN. This large dataset overcomes the problem of overfitting within the designed ANN, while extended training epochs avoid the underfitting problem. An extensive trial and error alternatives selection process was used to select the ANN optimum topography. The Nesterov-accelerated adaptive moment estimation algorithm is the optimization algorithm used to improve the ANN solution's training efficiency and convergence speed. The developed ANN achieved 93.09% and 92.62% accuracy for the apparent surface torque and the standpipe pressure feature, respectively, in the non-biased testing of the result. The work investigating the low-order topography for the ANN shows poor accuracy against the high and more sophisticated topography of the ANN. One of the ANN's behaviors realized is that enhancing the prediction accuracy for one feature results in a deterioration in the prediction accuracy of the other. Several attempts were made to create an automated drilling system; however, these attempts focused on the larger picture of the model and ignored the vital components that the calculated and predicted calculations are based on. System standpipe pressure and apparent surface torque prediction provide a solid foundation for an integrated system. The system's development used non-stochastic gradient decent tools to achieve the global minimum of the solution, contrary to most developed models' approaches to that topic. The high prediction accuracy of the developed ANN using the large dataset for training is a differentiator for this model.