Applying Machine Learning to Predict the Rate of Penetration for Geothermal Drilling Located in the Utah FORGE Site

Abstract

Well planning for every drilling project includes cost estimation. Maximizing the rate of penetration (ROP) reduces the time required for drilling, resulting in reducing the expenses required for the drilling budget. The empirical formulas developed to predict ROP have limited field applications. Since real-time drilling data acquisition and computing technologies have improved over the years, we implemented the data-driven approach for this purpose. We investigated the potential of machine learning and deep learning algorithms to predict the nonlinear behavior of the ROP. The well was drilled to confirm the geothermal reservoir characteristics for the FORGE site. After cleaning and preprocessing the data, we selected two models and optimized their hyperparameters. According to our findings, the random forest regressor and the artificial neural network predicted the behavior of our field ROP with a maximum absolute mean error of 3.98, corresponding to 19% of the ROP’s standard deviation. A tool was created to assist engineers in selecting the best drilling parameters that increase the ROP for future drilling tasks. The tool can be validated with an existing well from the same field to demonstrate its capability as an ROP predictive model.

Interpretable Feature Construction and Incremental Update Fine-Tuning Strategy for Prediction of Rate of Penetration

“…Intelligent models aim to approximate the complex relationship between the ROP and influencing factors by leveraging their powerful non-linear fitting capabilities. Artificial neural networks (ANNs) have been widely used, demonstrating promising results in ROP prediction [16,17]. Other machine-learning models, such as random forest [16,18,19], extreme gradient boosting [14,20], long and short-term memory (LSTM) networks [21,22] and hybrid networks [23,24], have been applied to predict ROP based on historical drilling data.…”

“…Artificial neural networks (ANNs) have been widely used, demonstrating promising results in ROP prediction [16,17]. Other machine-learning models, such as random forest [16,18,19], extreme gradient boosting [14,20], long and short-term memory (LSTM) networks [21,22] and hybrid networks [23,24], have been applied to predict ROP based on historical drilling data. Bizhani et al [25] addressed the issue of uncertainty in data-driven models by developing a Bayesian neural network model for predicting ROP.…”