Well log analysis data plays an important role in determining the reserves estimation of oil and gas, geophysical characteristics of reservoir, and wellbore stability. Sonic log is used to obtain the porosity along with the density and neutron log, and is also used to confirm the discontinuous rock formations or structures of the strata. However, there were frequent cases in which it was difficult to fully acquire the sonic log due to economic problems such as cost, defects in equipment, loss in the recording or transmission/reception stage, and wellbore problems. To solve this problem, studies were attempted to predict the sound wave detection using empirical correlation. However, the method using the empirical formula has limitations in its application in that the accuracy varies depending on a specific rock formation and geographic area.
Recently, various attempts have been made using machine learning and deep learning techniques. Among machine learning, random forest is an ensemble method that learns from multiple decision trees as an advanced technique of the decision trees. Decision trees have the advantage of having high explanatory power for the data, while have some problems of not having high predictive ability and poor accuracy due to overfitting. Random forest is a model that compensates for these shortcomings. Through bagging, data of the same size as the original data are arbitrarily extracted and generated, and decision trees are constructed based on this. It is a method to make a final prediction by combining observations through multiple decision tree models composed of randomly selected variables with a majority vote or average value. Random forests are being used for various problems such as classification and regression.
In this study, a predictive study was performed on the unmeasured sonic log of adjacent boreholes using well log data from Volve oil field provided to the public, The random forest model, which shows good performance through an ensemble of decision trees, was applied to model construction. The input variables to be trained in the model were selected through correlation analysis, and to evaluate the reliability of the well log prediction, SVDD (support vector data description) was used to classify and visualize the areas in terms of the prediction reliability. As a result, the methodology presented in this study was confirmed that it was highly useful as a method to identify high-reliability zones in well log prediction.