Training-while-drilling approach to inclination prediction in directional drilling utilizing recurrent neural networks

Abstract

Machine Learning adoption within drilling is often impaired by the necessity to train the model on data collected from wells analogous in lithology and equipment used to the well where the model is meant to be deployed. Lithology information is not always well documented and fast-paced development of drilling equipment complicates the challenge even further, as a model would likely become obsolete and inaccurate when new technologies are deployed. To bypass this problem a training-while-drilling method utilizing neural networks that are capable of modelling dynamic behaviour is proposed. It is a continuous learning approach where a data-driven model is developed while the well is being drilled, on data that is received as a continuous stream of information coming from various sensors. The novelty in presented approach is the use of Recurrent Neural Network elements to capture the dynamic behaviour present in data. Such model takes into account not only values of the adjacent data, but also patterns existing in the data series. Moreover, results are presented with a focus on the continuous learning aspect of the method, which was sparsely researched to date. A case study is presented where inclination data is predicted ahead of the inclination sensor in a directional drilling scenario. Our model architecture starts to provide accurate results after only 180 m of training data. Method, architecture, results, and benchmarking against classical approach are discussed; full dataset with complete source code is shared on GitHub.