The objective of this work is to present a first step towards a hybrid approach between machine learning (ML) and physics-based modelling to provide decision support for drilling problems. The motivation for developing a hybrid approach is to obtain methods that are more reliable and easier to automate than physics-based models, while still have enough accuracy and predictivity. In this first step, we replicate the performance for predicting downhole pressure in a well of a high-fidelity simulator based upon physical principles by using ML methods. In addition, we also suggest a future roadmap.

Methods, Procedures, Process

A high-fidelity physics-based model for drilling and well control operations is used to generate vast amounts of data for two cases, drilling with not major events, and drilling into an over-pressured reservoir. Key simulation input parameters and assumptions are varied to create realistic scenarios. We replicate the high-fidelity simulator downhole pressure predictions by two supervised machine learning algorithms. Random forest (RF) and recurrent neural network (RNN). The hybrid approach is flexible and is also employed for kick detection and estimation of the mass of the influx. After using unaltered data from the high-fidelity simulator, we also demonstrate the ML methods on corrupted data with synthetic noise.

Results, Observations, Conclusions

RF and RNN obtained very high accuracy, predicting bottom hole pressure with small error margin. Good results were also obtained for the kick estimation and kick detection cases. Tested on corrupted data, RF trained with noise performed significantly better compared to RF trained without noise, at the cost of a slight reduction in accuracy in the error free scenario. Initials tests on real data are ongoing and further work is needed. Hybrid methods have the potential of performing well with noisy environments and are valuable tool to be used in drilling automation.

Novel/Additive Information

Combining highly advanced dynamic models for drilling and well control with modern ML methods has not been done earlier to the best knowledge of the authors. Demonstrating this on real data will be valuable because data-driven and physics-based approaches used separately are considered inadequate for future automated drilling concepts.

You can access this article if you purchase or spend a download.