Exploration of Strategies to Improve Continual Learning From Irregular Sequential Drilling Data

Abstract

Abstract Economic and environmental objectives incentivize drilling wells faster, increasing the rate of penetration (ROP). A prerequisite for ROP optimization is an accurate ROP model that provides the expected rate given a set of drilling inputs and in response to downhole parameters in real-time. Relatively seldom and often different drilling operations make it difficult to collect a large dataset for supervised training of universal ROP predictors. Thus, ROP Deep Learning (DL) models often suffer from poor generalization. This work introduces continual learning as a new paradigm for data-driven ROP modelling. We subdivided historical drilling data from the Volve field and Marcellus Shale into smaller sections to imitate actual operations with irregular time-series data coming sequentially. In continual learning, each incoming section serves first as a testing set and is subsequently added to the training set. Using transfer learning, we continuously fine-tune the current DL model with varying amounts of recent data to adapt it to new conditions. Continual learning improved the ROP prediction accuracy for both J-shaped-well and horizontal-well test cases compared to the conventional one-off training paradigm. Furthermore, we documented the model performance for different fine-tuning configurations through a series of sensitivity analyses. We also discussed the challenges and trade-offs when implementing continuous learning for ROP prediction.