This paper introduces a new system to optimize rate of penetration (ROP) while drilling by applying statistical learning methods for ROP prediction and providing real-time performance increases and closed-loop control. This paper establishes the viability of such a method, and then goes on to present a framework for phased development and integration into drilling rig operations. By developing such a system, the authors hope to provide a tool to the oil industry which can improve drilling efficiency, reduce non-productive time (NPT), and lower the cost of drilling.
The statistical model at the core of this system relies on complex machine learning techniques like random forests, neural networks and ensemble techniques to build a predictive performance model of the well from both real-time data and data from previously drilled offset wells. Each rock formation type is analyzed separately, allowing the model to anticipate ROP performance changes at formation boundaries. This model can then assess the controllable drilling parameters and determine whether there is an opportunity to increase ROP by adjusting these parameters to some other value.
The necessary real-time drilling data can be accessed on the WITS/WITSML protocol, and a framework for reading these data into our model in real-time was developed. By using parameter selection techniques, parameters which are significant in predicting ROP were identified. Consequently, an advanced statistical model was developed to process the available data into ROP predictions in real-time, and estimate changes on surface controllable parameters such that the optimum ROP could be reached. A three-phase testing and integration path was developed for taking this tool from concept to completion. The first phase is to apply the statistical model to data from wells which have already been drilled and determine the amount of time which could have been saved by optimizing ROP. Then the second phase is to integrate the system with real-time data acquisition and provide a passive indicator of the current well's optimization level with a red-yellow-green traffic light style indicator. The final stage is to implement closed-loop control, allowing the model to run in a fully automated state. Throughout this implementation, it was shown that there was substantial potential for ROP improvement by following the suggestions of this statistical model.
This concept is innovative as it relies only on data which can be measured in real-time on the rig floor, and does not require information about rock strength, formation drillability, or other downhole properties. This paper provides the path for integrating this new statistical drilling optimization model into drilling operations, and provides a blueprint to bring this concept from simulation to the field.