Dynamic classification committee machine-based fluid typing method from wireline logs for tight sandstone gas reservoir

Tight sandstone fluid identification is difficult, and the intelligent algorithms can help to establish the fluid identification model of tight sandstone. Compared with the individual intelligent algorithm, the classification committee machine could improve the overall performance of the intelligent model by combining multiple experts (intelligent algorithms). For the limitation that the performance of individual expert is difficult to improve in the classification committee machine, adding a gate network to build the dynamic classification committee machine is a more effective approach to modular learning. In this study, firstly, the input data is divided into multiple subdatasets by using the gate network, and then the sub-datasets are trained by the decision tree, probabilistic neural network, Bayesian classifier, BP neural network, and nearest neighbor algorithm to obtain multiple submodels. Finally, the combiner is used to optimize the submodel combination to obtain the best fluid recognition model. Based on the logs and test data of the tight sandstone formation in the Dabei, Keshen and Bozi block of Kuqa Depression in Tarim Basin, the mean influence value method was used to obtain the sensitive logs data, and then a dynamic fluid typing model was constructed. The training and verification accuracy was 96. 29% and 91. 39%. Taking well BZ9 as an example, the fluid type was predicted, and the prediction result is agreement with the test result. This method combines unsupervised and supervised learning, introduces the gate network to improve the dataset utilization, and avoids the impact of uneven data distribution on model training. The voting strategy is used to integrate multiple experts and establish the correspondence between the submodels and the experts. The prediction accuracy and generalization of the fluid identification model are greatly improved.

Automated summary

Tight sandstone fluid identification is challenging, but intelligent algorithms can help establish a model. Using a classification committee machine with a gate network for modular learning improves overall performance. The method enhances prediction accuracy and generalization by combining multiple experts and establishing correspondence between submodels and experts.