Projection of Logging While Drilling Data at the Bit by Implementing Supervised Machine Learning Algorithm

Abstract

Abstract To analyze drilling performance a combination of Logging While Drilling data (LWD) and surface drilling data is combined. However, distance between some of the sensors, and the bit is greater than 20-30m (66-98 ft). In this case, determination of the LWD data at the bit becomes essential. This paper aims to implement machine learning algorithms to predict LWD data at the bit. The results of the model can be used to perform real-time analysis that considers the alterations in petrophysical properties, lithologies and rock strengths while drilling, without the drawbacks of LWD sensor offset. The aim of the paper is to predict LWD data at the bit by evaluating which supervised machine learning algorithm to incorporate. For training and validation of the model, a dataset of high porosity formations from multiple wells located in the North Sea has been used. Dataset included gamma ray (GR) log data recorded near the bit and drilling parameters recorded at the bit. Multi-linear regression (MLR), K-nearest neighbor (KNN) regression, random forest (RF) regression and support vector machine (SVM) regression are used for model building. The most efficient model with the best coefficient of determination (R2) is selected. The prediction forecasting for the random forest regression model was better among all the previously discussed regression models. The R2 value for the random forest regression model 98% and the KNN regression model came in second with R2 value at 95%. The worst performing regression model was the multi-linear regression model. This machine learning approach to consider the LWD sensor offset can be useful in the determination of petrophysical properties at the bit and in the real-time drilling analysis.