Machine Learning Techniques Application for Real-Time Drilling Hydraulic Optimization

Abstract

Abstract Over the past decade, several methods and techniques have been proposed to optimize drilling hydraulic's in real-time; one of these techniques is machine learning, which has shown promising results in prediction and optimization. Nevertheless, the real-time implementation of these techniques is still challenging since most of the published work tried to perform prediction tasks rather than the optimization task. In this regard, this paper tries to tackle the shortcomings of the recently published related methods by presenting a holistic model, based on a machine learning concept, focused on real-time optimization of drilling hydraulic's within a sufficiently short time span and without disturbing the drilling process. The presented approach relies on using two interconnected models to achieve the goal, which can be classified into, data-driven and analytical models. The real-time optimization process starts by using two predictive models to predict standpipe pressure and annular pressure losses and an analytical model to compute the drill-string pressure loss. Subsequently, the three generated values are used by an optimizer algorithm to generate the optimum combinations of surface drilling parameters, namely, weight on bit, flow rate, and rotation per minute, which are expected to optimize drilling hydraulic. Two case studies were conducted based on a historical drilling data set to assess the performance of the utilized predictive models and to measure the time required for the model to perform an optimization task. The results reveal that the predictive model demonstrated very high accuracy in terms of predicting SPP and APL as indicated by the determination coefficient value (R2), which was between 0.87 and 0.99. Moreover, the overall simulation time was within a range of between 2 to 4 minutes, which is considered a rational time frame for a real-time optimization task. The methodology applied allowed us to conclude, even showing some limitations, that machine learning techniques can be well used for hydraulic optimization in real-time.