Comparative Analysis of Machine Learning Based Feature Selection Approach for Carbonate Reservoir Cementation Factor Prediction

Abstract

Abstract A key component of the fourth industrial revolution is data integration. However, this comes with a major challenge: handling increased input feature dimensionality. Multivariate feature space increases model complexity, memory utilization, and computational intensity, thereby reducing model performance. A pragmatic approach to input feature space reduction is therefore required. This paper presents a comparative study of the performance of a nonlinear feature selection methodology based on fuzzy ranking (FR). The FR algorithm is extracted from a segment of Fuzzy Logic, an existing machine learning technique. The performance of this feature selection algorithm is tested and validated with respect to the prediction of cementation factor as a log from wireline measurements using machine learning techniques. Cementation factor is denoted by the exponent m in Archie's equation. A subset of the log data selected by the FR algorithm is automatically fed into artificial neural network (ANN) and support vector machine (SVM) models to build FR-ANN and FR-SVM hybrid learning models. A multivariate linear regression (MLR) model is also implemented. The performance of the hybrid models is compared to those of MLR, ANN and SVM without the feature selection procedure. We further compare the outcome with ANN and SVM fed with linearly correlated input features. The hybrid learning methodology is driven by patterns discovered in the data and eliminates subjective human bias in the choice of the input features. It also takes into consideration the possible nonlinear relationship between the wireline logs and m. The FR-ANN model shows improved performance over the other models with the highest correlation coefficient and lowest root mean squared error. The performance of the FR-ANN hybrid model demonstrates the efficiency of the proposed nonlinear feature selection hybrid methodology. A future work will apply this approach to high dimensional, integrated data types from many wells. We expect that the outcome will significantly improve the prediction accuracy and further impact reservoir models using the predicted properties.