Predictive data-driven analytics has driven massive interest, primarily because of its successful implementation in several industries. Data-driven modeling and its application to predict downhole environment will prove to be the future of drilling operations, as they hold the potential to optimize highly complex drilling operations. In general, rate of penetration (ROP) optimization involves adjustment of the weight on bit (WOB) and rotary speed (RPM) for efficient drilling. ROP follows a complex relationship with several other parameters such as formation properties, mud properties, mud hydraulics, borehole deviation as well as the size/type of bit.
Traditional regression analysis models have limitations and have limited accuracy while attempting to describe the dependence of one observed quantity on another observed quantity. On the other hand, the artificial intelligence methods face drawbacks trying to understand the physics behind the operations. To ensure the physical and technical feasibility of the prediction, coupling conditions between the two have been developed for the ROP optimization.
In this study, a new model based on statistical regression and artificial neural networks (ANNs) was designed to predict ROP using field data gathered from the North Sea Horizontal Wells. Exploratory analysis was performed to find correlations between variables, followed by measurement of predictor importance to infer relative contributions and weights of inputs. Cross-validation was employed to prevent overtraining of models. Several models such as Step-Wise regression, neural networks (NN, KNN), support vector regression (SVR), classification regression trees (CART), were applied for prediction. Ensemble methods such as Random Forests (RF) and Boosting helped increase accuracy and reduce errors. The algorithm designed was further tested on other wells and was shown to predict with significant accuracy. This proves that several parameters need to be comprehensively considered while optimizing the ROP. Since the algorithm presented here doesn't depend on a single model but rather several predictor models, it can be effectively employed independent of location or formation.