Machine Learning Vs. Statistical Models Approach: A Case for Probabilistic AFE Cost in Drilling

Abstract

Abstract Traditionally, statistical AFE (Authorization For Expenditure) cost considers the probabilistic variance of time which is can be obtained by carefully using historical data. However, the data pattern may not be well captured to represent a realistic variance of drilling duration due to the nature of different operations complexity. This paper describes the development and comparison of machine learning and statistical models as well as their integration to estimate the drilling duration for probabilistic AFE cost prediction. Exploratory data analysis and data pattern recognition are performed to model the well's parameters that affect the well drilling duration. Subsequently, the random forest (RF) and multiple linear regression (MLR) models are developed to estimate the well duration for any well complexity. To capture the variation of the input data, Monte Carlo (MC) simulation is then employed to construct 100 simulation cases prepared using the developed RF and MLR models. Finally, the probabilistic variance of drilling duration from these two models are fed into the data-driven models to predict AFE cost. Based on blind-test results, it is observed that the combination of RF and MC simulation models provides a more realistic variance of drilling duration compared to other models. It can reduce the uncertainty provided by a purely statistical model which is a combination of MLR and MC simulation models. In addition, it also provides a better accuracy against the standalone RF model. Incorporating the probabilistic variance of time from a combined RF and MC models into the random forest structure for well cost prediction resulted in better model performance compared to other data-driven-based well cost models developed using multiple linear regression, decision tree, and artificial neural network algorithms. In terms of the R-square value between predicted and actual costs, the AFE accuracy can be improved from 0.74 to 0.91 using the proposed model. The cumulative density function of the simulated time and cost provides insight to evaluate the drilling duration and AFE cost based on P10, P50, and P90 values. Tens of cases are performed and the support vector machine (SVM) classification plot of the probabilistic AFE cost against historical wells data is generated indicating the range of possible cost clusters. The proposed methodology enables drilling industry personnel to quickly estimate the probabilistic time and cost of the project during well planning and during drilling using widely available drilling variables. This novel approach provides a realistic and efficient alternative to explore hundreds of scenarios encompassing the range of drilling variables uncertainties in predicting the probabilistic AFE cost.