Effectiveness of Hybrid Algorithms for Optimized Sub-surface Property Prediction Over Conventional Geostatistics
Abstract
The use of spatial interpolation methods can be attributed to the non-uniform distribution of the geological properties in the subsurface. The limited success has stemmed out of the fact that these methods made us appreciate and evaluate the sub-surface resources in 3D in terms of its characterization and its commercial potential. Common geostatistical methods such as Ordinary and Universal Kriging, and Inverse Distance Weighted Interpolation (IDW), have been the main stay of the industry for predicting reservoir properties in the sub-surface. Despite their wide acceptance, associated uncertainties lead to the missing exploration and development target, particularly while handling complex strati-structural situations. Interpolation in heterogeneous geological conditions using Kriging is computationally expensive and IDW, though simple, does not provide accurate estimates. Therefore, the prediction of properties in un-cored reservoirs pose a challenge to reservoir description and dynamics. With the development of statistical tools such as machine learning (ML) regression models, the ensemble modelling promised more accurate estimations in lesser computational time for reservoir property prediction. This work presents a comparative study between the conventional interpolation methods and the state-of-art Machine Learning techniques for three-dimensional reservoir property estimations using well-logs. In the first approach, we modelled different variogram models with varying parameters are analysed to find the parameters that best correlate with the input data. For the second approach, different Machine Learning techniques, including Random Forest, Boosting and Support Vector Machine, and their weighted combinations were investigated. Different Machine Learning and AI techniques using the parameters extracted from the variogram analysis. Running multiple realisations of the property volume, we demonstrated that the hybrid approach of Random Forest Regressor with three-dimensional Ordinary Kriging provided the most accurate estimates in the validation data. The results show that hybrid approaches are novel and highly accurate for prediction with error significantly low compared to just the Machine Learning algorithms alone with the same time complexity.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFMIN45A..06S