A comprehensive machine learning model for lithology identification while drilling

Abstract

A comprehensive machine learning model is established for lithology identification while drilling. Easily available field data, including well trajectory parameters and real-time drilling parameters, are chosen as feature parameters. First, kernel principal component analysis is used to convert the original feature space, enhancing the association between feature parameters and corresponding lithology. Subsequently, clustering is performed in a novel way to magnify the differences between data points, which ameliorates the ability of the model to resist signal noise and reduces the difficulty of lithology identification. Finally, for each data cluster, multiple base classifiers are integrated to constitute strong sub-classifiers through the Stacking method, which improves the classification accuracy. The model is validated with field data of a shale gas reservoir in East Sichuan. With the optimal model parameters, the overall identification accuracy of the comprehensive machine learning model is 0.87. The specific identification accuracies are 0.88 for sandstone, 0.85 for mudstone, 0.86 for limestone, and 0.89 for shale. The current comprehensive machine learning model is also demonstrated to have better performance than the single-classifier models and the deep neural network model. Real-time field application of the model during the drilling process of a new well shows an accuracy of 0.92, and the rate of penetration is increased by 22.74% through the matching of drilling parameters and formation lithology. It is proved that the present comprehensive machine learning model can be employed for lithology identification while drilling, and provide guidance for adjusting drilling parameters in real time.