Predicting Rate of Penetration Using Artificial Intelligence Techniques

Abstract

Abstract Rate of Penetration referrers to the speed of breaking the rock under the bit. It measures the speed or the progress of the bit when it drills the formation. It has been reported in the industry that high percentage of the well budget is spent on the drilling phase, thus many drilling operators pay close attention to this factor and try to optimize it as much as possible. However, it is very challenging to capture the effect of each individual parameter since most of them are interconnected, and changing one parameter affects the other. As a result, many companies maintain a data for the drilling performance per field and set certain benchmarks to gauge the speed of any newly drilled well. To date, no solid or reliable model exists because of the complexity of the drilling process, and one cannot capture every factor to predict the rate of penetration. Therefore, the utilization of artificial intelligence (AI) in the drilling applications will be a game changer since most of the unknown parameters are accounted for during the modeling or training process. The objective of this paper is to develop a rate of penetration model using artificial neural network (ANN) with the least possible number of inputs. Actual field data of more than 4,500 data points were used to build the model. The inputs were pumping rate, weight on bit, rotation speed, torque, stand pipe pressure and unconfined compressive strength. Well-A was used to train and test the model with 70/30 data ratio. Then two unseen data which are well-B and well-C were used to test the model. ANN was used in this study, with many sensitivity analyses to achieve the best combination of parameters. The obtained results showed that ANN can be used effectively to predict the rate of penetration with average correlation coefficient of 0.94 and average absolute percentage error of 8.6%, which shows 22% improvement over the current published methods. The best ANN model was achieved using 1 layer, 12 neurons and a linear transfer function. The developed ANN-ROP model proved to be successful using only six inputs and having a total of two wells with unseen data.