A Comparison of Machine Learning Algorithms for Rate of Penetration Prediction for Directional Wells

Abstract

AbstractOne of the biggest challenges facing oil and gas companies is to lower the cost of drilling operations. The most critical parameter affecting drilling cost is the rate of penetration (ROP). Improving the ROP and reducing the drilling cycle can be significant for companies to reduce drilling costs and risks, thus enhancing market competitiveness. In the present study, we evaluate the accuracy and effectiveness of different machine learning algorithms for directional wells. We collect many field drilling datasets such as bit type, bit drilling time, revolutions per minute (RPM), weight on bit, torque, formation type, rock properties, and hydraulic and drilling mud properties. We input these data to the machine learning model to be trained, validated, and tested for predicting ROP. The machine learning models we used include linear regression, Artificial Neural Networks (ANN), Support Vector Machine (SVM), and Random Forest (RF) algorithms. In this study, we implement the Genetic Algorithm (GA) to optimize the hyper-parameters of the five machine learning algorithms. We also apply Savitzky-Golay (SG) smoothing filter to reduce the noise in the original dataset. We use accuracy metrics such as root mean square error, mean absolute error, and regression coefficient (R2) to compare the accuracy of several machine learning models. At last, we select the best-performing algorithm as the prediction tool for ROP. We conduct 50 cases for each machine learning model, where we evaluate the performance of the models and measure the time required for the models to be trained for the prediction tasks. The comparative study shows that implementing the GA optimization algorithm increased the accuracy of individual ROP models. We find that optimizing only a few hyper-parameters can significantly improve the machine models’ accuracy. We also compare the results from the model trained by the data processed with the SG smoothing filter with those trained by the original data. The study demonstrates that the SG algorithm effectively improves the accuracy of the machine learning models. Among the four machine learning models, ANN has the highest accuracy after GA optimization, reaching 97% on average. The overall training time for all four algorithms is between two and four minutes, considered a reasonable time frame for a real-time training and prediction task. We compare several machine learning methods’ accuracies in predicting ROP in real-time. We find that ML-based prediction models, especially ANN with hyper-parameters optimized by Genetic Algorithm, can accurately predict ROP in real-time and provide the operator with suggestions for appropriate measures.