Estimation of Lost Circulation Zones Using Random Forest Classification and Image Logs

Lost circulation is one of the most chronic and most expensive problems in the drilling process as it leads to enormous losses in terms of finances and time. The conventional prediction methods mainly are based on the traditional well logs and correlations which have been empirically determined and which fully exploit the rich textual information available in the Formation Microresistivity Imaging (FMI) logs. The work is the innovative supervised machine learning model based on the use of the Random Forest classification algorithms to estimate the lost circulation areas based on the combination of the traditional Logging While Drilling (LWD) data with the texture features of the FMI image logs. The methodology was used on a large dataset of six wells of the Niger Delta, and this includes more than lost circulation events recorded by drilling reports. The conventional logs (gamma ray, resistivity, bulk density, caliper) were used as the input parameters complemented with the texture descriptors based on the FMI images determined by the Gray Level Co-occurrence Matrix (GLCM) methods. GLCM texture characteristics such as contrast, correlation, homogeneity and energy were calculated with the help of Open CV and at 0, 45, 90, and 135 angular orientations. Preprocessing of data included removing outliers, lithological coding, and depth matching of traditional logs with image-based characteristics. Random Forest classifier was trained on 70% of the data and was validated on the 30% of the data with 100 estimators and the maximum depth of 10 implemented in scikit-learn with Python. Accuracy, precision, recall, F1-score and ROC-AUC were used to measure model performance. Findings indicate that the combined method had an overall prediction accuracy of 86% and the image-based texture characterizations increased accuracy by 14% points as opposed to the traditional log-only models (71% to 86%). The ROC-AUC rose by 0.79 in case of conventional features and 0.91 of the combined feature set. The analysis of the feature importance showed that the most important features were texture contrast and caliper measurements. The methodology will allow the identification of lost circulation areas at an initial stage during the real-time implementation of drilling activities and allows taking proactive measures to mitigate the latter and improve the productivity of operations.