Log identification of fluid types in tight sandstone reservoirs using an improved Stacking algorithm
Tight sandstone reservoirs have poor physical properties.Resistivity logging data usually reflect the lithological and physical properties of the formation,and the logging response characteristics caused by formation fluids are insensitive.Therefore,the application of traditional well logging interpretation charts to classify reservoir fluids has low accuracy and cannot accurately classify fluid types.Machine learning can input more dimensional features for modeling,and can find more potentially effective information than traditional well logging interpretation charts.At present,machine learning algorithms mainly consist of a single algorithm and an ensemble learning method.Compared with individual machine learning algorithms,ensemble learning combines the training or prediction results of a single algorithm through reasonable decision-making,and finally outputs the final result.However,the performance gap between different ensemble learning strategies is large,so the structure of the ensemble learning model needs to be improved.This paper proposes an improved Stacking algorithm,which uses the Mean Impact Value(MIV)method to find logging curves that are more sensitive to fluids for input.It studies the use of different feature sets to build multiple sub-models,and uses different integration strategies to combine the sub-models into Expert models,these expert models have better performance than sub-models.In order to avoid overfitting,we designed an independent expert model,simulated the prediction results of the expert model through cross-validation,and input the simulated prediction results into the meta-classifier for learning to obtain the final prediction results.We applied the re-improved stacking algorithm to identify fluids in the tight sandstone reservoir of the Dibei area in the Kuqa Depression.For comparison,we input the same logging data to the CatBoost model and the XGBoost model.The results show that the improved stacking algorithm achieved the highest accuracy,up to 93％,which is better than the accuracy of the comparison model.The above research proves the effectiveness and applicability of the improved stacking algorithm,and provides a new idea for intelligent methods to identify fluids in tight sandstone.
万方数据
维普
