Developing a new rigorous drilling rate prediction model using a machine learning technique

Drilling rate of penetration (ROP) prediction is an enormously important step to optimize drilling controllable parameters. Therefore, numerous efforts have been done in order to present a more precise estimator model that is still ongoing. The results of the literature review indicated that, between the two approaches followed for ROP modeling, namely physic-based models and data-driven methods, the use of the data-driven methods has grown significantly. The literature review made two points clear: (1) the geomechanical parameters were not considered adequately, and (2) no previous research had applied a combination of least-squares support-vector machines (LSSVM) with cuckoo optimization algorithm (COA), particle swarm optimization (PSO), and genetic algorithms (GA). This led us to use hybrid algorithms for ROP modeling at vertical wells drilled in the southwestern Iran. For this purpose, mud logging parameters (including depth (Depth), mud density (MD), torque (Tor), standpipe pressure (SPP), equivalent circulating density (ECD), weight on bit (WOB), revolutions per minute (RPM), flow rate (FR), and ROP) and geomechanical parameters (including gamma ray (GR), porosity (NPHI), density (RHOB), and uniaxial compressive strength (UCS)) were collected along the studied wells. Histogram analysis and pairwise evaluation of the studied features indicated the presence of outliers among the data points. Accordingly, the Tukey's method was used to omit the outliers. Subsequently, the most relevant features in the ROP prediction were selected using a combination of the non-dominated sorting genetic algorithm II (NSGA-II) with multilayer perceptron (MLP) neural network. The results showed that one could attenuate the modeling error by increasing the number of input parameters into the ROP estimation model. However, the improvement was very subtle when the number of input parameters exceeded six. Therefore, six parameters (UCS, FR, WOB, Depth, MD, and RPM) were used for ROP modeling using LSSVM-COA, LSSVM-PSO, and LSSVM-GA hybrid algorithms. Results of applying these hybrid algorithms indicated that the LSSVM-COA outperformed the other two algorithms in terms of accuracy and the coefficient of determination. In a next step, the LSSVM-COA was trained on both the outlier-omitted and raw datasets, indicating the important role of the outlier omission step in achieving a highly accurate and reliable model. Continuing with the work, in order to evaluate the performance of