Unconventional reservoirs comprise of various heterogeneous productive and non-productive units which can be correlated with facies. To focus a target zone during drilling, it is essential to understand and identify unique zones in real-time. However, real-time LWD/MWD tools provide formation properties data with depth and time delay. Machine learning (ML) can help in predicting productive/non-productive facies/rock types without any time and depth delay enabling early decisions resulting in optimization of rig time and cost. In this study, ML approach has been employed to predict the frackable facies in a horizontal well in an unconventional reservoir using real-time surface drilling parameters and formation stress properties.
Initially, the whole data was pre-processed by visualizing through matrix scatterplots and histograms and outliers were removed. Optimum number of clusters were estimated using sum of squares within (SSW) and Silhouttee techniques. MinMax scaling methodology was used to scale up the elastic properties before clustering/labelling. Drilling data was also scaled in the supervised learning before predicting the labels. In the unsupervised learning, the data was labelled through K-means Clustering with 3 number of clusters. The supervised learning techniques used were 1) K-Nearest Neighbors (KNNs) classifier; 2) Support Vector Machine (SVM) classifier; and 3) Random Forest (RF) classifier. 80% of drilling data was used to train the classifier, whereas 20% of the data used to test the ML classifier. The hyperparameter grid optimization with 10 cross fold validations was also performed to optimize the parameters of all the classifiers. The variable importance was also evaluated through RF classifier to analyze the impact of sensitive drilling parameters on the predicted facies. Furthermore, confusion matrix and accuracy score of the best parameters also obtained and compared.
This study showed that KNNs, SVM, and RF classifiers predicted frackable facies with 78%, 78.5% and 76.6% accuracy respectively on 20% test data set. These results are based on hyperparameter grid optimization. Also, facie (2) was found to have the highest brittleness index where facie (3) was found to have the highest ductility. The accuracy of KNNs, SVM and RF classifiers based on confusion matrix was 79%, 74% and 77% respectively. The KNNs classifier outperformed both SVM and RF based on testing data. Additionally, it was observed that gamma radiation at bit and standpipe pressure were the most critical parameters in relating to confusion matrix accuracy and had variable importance of ~30% each. The supervised learning algorithms predicted lithology with over 75% accuracy, showing the robustness of data-driven modeling approach.
The success of ML based modeling approach can improve real-time decision making. This data-driven predictive model can also be extended to other wells in conventional rock formations.
