An optimized ensemble for predicting reservoir rock properties in petroleum industry

The estimation of initial hydrocarbon in place before investing in development and production is the main objective in petroleum industry. Porosity, permeability and water saturation are the most important key variables to quantitatively describe petroleum reservoir. However, identification of these parameters which relies on core data analyses is expensive and time consuming. A lot of researches have been done to predict the reservoir parameters using well log data through applying various methods. To predict theaforementioned parameters, we need a method with high accuracy,good generalization, fast and low in cost. In the present thesis, we proposed a new method named optimized ensembleto improve the prediction of these reservoirs parameters from well log data with the aid of available core data. Ensemble is a learning algorithm that combines some experts instead of considering a single best expert for the predictions.The thesis proposed anoptimizing method leading to small structure of assemble GA. After constructing suitable ensemble members, we need to combine them with a propermethod to improve the accuracy.So, we proposed two combining methods to improve the prediction accuracy while maintaining the generalization. The first method isbased on fuzzy genetic algorithm to overcome the premature convergence. The second method is based on two other functions instead of traditional fitness function in genetic algorithmnamely MSE to determine the individual's weight in an ensemble.This approach is based on Huber and Bisquare functions which are meant to avoid the influence of outliers that can be found in many real data such as geosciences data. In the present thesis, we implemented our method for predicting these three most important reservoir parameters namely porosity, permeability and water saturation.The real field data is obtained from Iranian offshore and onshore oil fields. A total of 3695 data points from the 5 wells having conventional well log data and core data were used. Threeperformance measurements for analysing and comparing the predicted results and target values including correlation coefficient (R),Root Mean Squared Error (RMSE) and related RMSE were selected. The results on pruning method show that the memory requirements for porosity,permeability and water saturation decreased to 68.75, 68.75 and 81.25 percent respectively.The results on pruned ensemble with FGA based weighted averaging also show that triple performance measure (RMSE, RRMSE, R/R2) improved (9.95, 12.50, 1.16) percentfor porosity, (6.6, 16.21, 1.17) percentfor permeability and (37.56, 28.08, 1.52) percentfor water saturation in comparison to the whole ensemble.A comparison results between the Huber and MSE based GA show that that triple performance measure (RMSE, RRMSE, R2) improved (17.3, 25.2, 1.0) percent for the permeability data set.

Preview PDF FK 2013 21RR.pdf Download (1MB) | Preview

Actions (login required)

View Item