Prediction of penetration rate in drilling operations: a comparative study of three neural network forecast methods

Optimizing purposes of the drilling process include reduction in time, saving costs, and increasing efficiency, which requires optimization of controllable variables and variables affecting the drilling process. Drilling optimization is directly related to maximizing the rate of penetration (ROP). However, estimation of ROP is difficult due to the complexity of the relationship between the variables affecting the drilling process. The main goal of this study is to develop three computational intelligence (CI)-based models including multilayer perceptron neural network optimized by backpropagation algorithm (BP-MLPNN), cascade-forward neural network optimized by backpropagation algorithm, and radial basis function neural network optimized by biogeography-based optimization algorithm (BBO-RBFNN) to estimate ROP. Also, in order to broaden the comparisons, some conventional ROP models from the literature were employed. The required data were collected from the well log unit and the final drilling reports of four drilled wells in two different oil fields in southwestern Iran. Firstly, all data were preprocessed to remove outliers; then the overall noises of the data were reduced by implementing Savitzky-Golay smoothing filter. In the next stage, nine input variables were selected during a feature selection step by combining the BP-MLPNN and NSGA-II algorithm. The results of this study showed that developed CI-based models more accurate than conventional ROP models. Also, a survey of statistical indices and graphical error tools proved that BBO-RBFNN model has the highest performance to predict ROP with values of APRE, AAPRE, RMSE and R2 equal to - 0.603, 5.531, 0.490 and 0.948, respectively.

Automated summary

The study aims to develop three computational intelligence-based models to estimate the rate of penetration, which are proven to be more accurate than conventional models through data preprocessing and feature selection.