Abstract
This paper presents the development of machine learning model to predict key performance indicators (KPIs) in oil and gas Well drilling, including Dry Hole Drilling Days (DHDD), Dry Hole Drilling Cost (DHDC), and Clean Time (CT). It explores diverse input variables to improve accuracy and reliability of these KPI predictions.
This paper employs a collaborative approach between academia and industry to develop machine learning models for KPI prediction using CRISP-DM (Cross Industry Standard Process for Data Mining) methodology. Methodology involves compiling and preparing the historical data from previously drilled Wells from multiple fields of geological complex and tectonically stressed region of Pakistan. Subsequently various machine learning algorithms are trained, tested and evaluated for predicting three drilling KPIs. Based on top performing machine learning models, a calculator is prepared and deployed in to order validate the results against the unseen data.
The results demonstrate the effectiveness of machine learning models in predicting drilling KPIs; Dry Hole Drilling Days (DHDD), Dry Hole Drilling Cost (DHDC) and Clean Time (CT) to an acceptable level of accuracy. Among the models assessed, Support Vector Machine (SVM), Random Forest (RF), and Stacking employing SVM as the base model and Linear Regression (LR) as the meta-model emerged as the top performing models. SVM demonstrated superior predictive capabilities for DHDD, achieving an Error Percentage of 14.2% on testing data. Whereas, on similar data RF excelled in forecasting DHDC with an Error Percentage of 11.2% and Stacking proved best for CT prediction with an Error Percentage of 10.6%. KPI predictions from machine learning model based calculator for the future planned Wells are compared with in-house cost and time estimation methods and found both methods complementing each other. These findings emphasize the potential of machine learning in optimizing drilling operations and maximizing efficiency. While a single machine learning model for all KPIs is not possible, the accurate prediction of DHDD, DHDC, and CT suggests the viability of implementing machine learning models for better budget planning, optimizing the drilling operations, and ultimately reducing the cost of drilling operations.
Machine learning model can accurately predict critical macro-level KPIs of drilling operations, such as DHDD, DHDC and CT. Larger number of machine learning algorithms are explored for this purpose. Previous application of machine learning models in the field of drilling is predominantly focusing on micro-level KPIs like drilling parameters optimization and Rate of Penetration (ROP) improvement. Whereas, this research represents a significant departure from existing literature by exploring the prediction of macro-level Well drilling KPIs.
