A Novel Preprocessing Method Based on Variational Mode Decomposition for Reservoir Characterization Using Support Vector Regression

Abstract

This article proposes a variational mode decomposition (VMD)-based preprocessing scheme to regularize a target porosity signal to improve its prediction from seismic attributes. A workflow consisting of preprocessing, prediction, and postprocessing stages is designed to implement the proposed scheme on a real hydrocarbon field dataset. In the preprocessing stage, the original porosity signal is decomposed into multiple intrinsic mode functions (IMFs). The regularized porosity signal is constructed from the decomposed IMFs by removing some of the higher frequency modes. Three alternate methods—empirical mode decomposition, complementary ensemble empirical mode decomposition, and VMD—are used for decomposition to identify the best regularization technique. In each case, normalized mutual information and entropy are used to regulate the filtering. Then, mapping of the porosity signal and seismic data is carried out by support vector regression (SVR). Four metrics—correlation coefficient, root-mean-square error, absolute error mean, and scatter index—are used to quantify the performance of SVR while using original and regularized porosity signals as targets separately. The tuned SVR parameters are then used to predict synthetic porosity logs over the study area from seismic volumes. Finally, the predicted porosity volume is postprocessed using an improved complex adaptive diffusion filter. The performance of the SVR is also compared with an artificial neural network. This framework can be used to predict different lithological properties from seismic attributes and visualize them over a reservoir area.