Evaluating the Rate of Penetration With Deep-Learning Predictive Models

Abstract

This paper presents a sophisticated deep-learning framework designed for predicting rate of penetration (ROP) by assimilating well-log data, litho-facies classifications, and parameters of onshore production wells drilling operations in Central Asia. The evolution in bit technology and relevant drilling operation underscores the necessity for enhancing the traditional empirically derived predictions. Distinctively, our approach integrates transfer learning into a conventional deep-neural-network, employing two important techniques. One is data quality control by Kalman filter to make machine learning applicable to in situ data which have significant noises. The other is K-means clustering to reflect litho-facies attributes as input features of deep-learning predictive model. The developed scheme was applied to the in situ drilling data which have 12 kinds of data types: measured depth; two drilling operation variables, namely weight on bit (WOB) and rotary speed (RPM [revolutions per minute]); six well-log measurements including density (RHOZ), neutron porosity (TNPH), resistivity (RT), sonic (DT), gamma ray (GR), and photoelectric factor (PEFZ); alongside three clusters delineating litho-facies data. The developed schemes are tested by being applied to the in situ well’s ROP prediction based on the training and validation of four wells’ data. All in-situ data are in the interval of 7-in. casing which ranges from about 800 to 3100 m. By adding the well-log-data-driven litho-facies and the transfer learning on the base model, ROP prediction performances are improved as follows: R² value up to 49% (from 0.49 to 0.73), mean absolute error up to 23% (from 6.79 to 8.82 m/h), and the dynamic time warping up to 24% (from 361 to 473 h), respectively. As a result of deriving a drilling operation strategy that allocates WOB from 1 to 6 tons for each 100 m section and optimizes ROP, it is expected to reduce drilling time by about 16.5% compared to actual drilling. The developed method can evaluate ROP with high reliability from the comparison between ROPs predicted and measured in actual drilling operation. It is expected that the developed scheme can be applied for an extension to real-time ROP optimization, a kind of inverse modeling, to find the optimum parameter conditions for ROP maximization, as a forward model.