Well Logs Estimation Using Multimodel Strategies and Feature Expansion Based on Seasonal Decomposition

Abstract

Geological properties deriving from well log records are key data for understanding reservoirs, quantifying in-place volumes, production potential and hydrocarbon reserves. However, log data acquisition is expensive and time consuming, often resulting in missing or incomplete information. This study aims to develop machine learning models to create synthetic porosity, compressional slowness, resistivity and bulk density logs, focusing on two new approaches based on seasonal decomposition to enhance model quality compared to usual supervised learning. Feature decomposition, expanding the input space in trend and combined seasonal-residual terms, was the winning strategy for three of the four evaluated logs and promoted a gain of 8.5% in R2 and reduced RMSE by 27% in the case of the compressional slowness log. Furthermore, the presented ensemble-based method, in which estimators are specialized in trend and combined seasonal-residual predictors, outperformed the baseline models for 6 of the 8 examined estimators for both porosity, compressional slowness and resistivity predictions. In addition, this strategy proved to be the best for the bulk density log, improving R2 by 1.7% and reducing RMSE by 5.2%. In all four scenarios examined, both methods achieved better results than the baseline strategy and proved to be robust techniques for well-log estimation purposes. Actual data from wells in the Volve oil field located in the North Sea were utilized to examine these methods.