Prediction, interpolation and extrapolation of drilling data with Deep Learning

Abstract

Directional drilling is an established technology within the petroleum industry. In this traditionally conservative and risk averse environment application of artificial intelligence encounters difficulties at various stages of the process. Additionally, the unique nature of drilling data makes deployment of the off-the-shelf algorithms problematic on multiple levels. An artificial neural network with a custom architecture is explored in this dissertation; it is combining recurrent elements as well as traditional artificial neurons to fully utilize information in both the dynamic behaviour of the system, as well as data that is available in real time. This fit for purpose algorithm unlocks significant improvements to the traditional directional drilling technologies. To complement this novel architecture, research is presented that focuses on the data preparation. Methods are presented that tackle characteristic problems of the real-time drilling logs that make the incompatible data digestible for the machine learning. New algorithms were developed that allow to gauge the difference between the raw and the processed data. Simplifying the field deployment a method for on-the-job training is explored for the developed architecture, where no prior knowledge about the exact drilling system nor historical data is required. To aid the confidence in the presented methods a fit for purpose sensitivity analysis is investigated allowing to peek inside the data driven algorithm.