Gate recurrent unit network combines with Adaboost algorithm to classify fluid types by well logging parameters

Abstract

Given the diverse range of fluid types in reservoirs, their frequent alternation, and complex composition, traditional methods exhibit low accuracy in identifying these types. To address this, we introduce machine learning techniques to predict fluid types by extracting logging data. However, a single Gate Recurrent Unit (GRU) network is insufficient to meet the demands of fluid type prediction. Therefore, we propose a method that combines the GRU network with the Adaboost algorithm, referred to as GRU-Adaboost. The GRU-Adaboost model effectively combines multiple weak classifiers into a more powerful classifier through iterative training and gradual adjustment of sample weights. By using a voting strategy to synthesize the predictions of individual classifiers, the impact of errors from each classifier can be reduced. Compared with traditional GRU networks and Long Short-Term Memory models, the proposed GRU-Adaboost model shows improved accuracy. To validate the feasibility of our method, we apply the proposed algorithm to three wells. Experimental results confirm that the prediction performance of GRU-Adaboost surpasses that of other models.