Improved recurrent neural network for complex lithology identification in gas reservoirs
Shaoqun Dong, Leting Wang, Tao Xu, Xu Bai, Xiangjun Wang, Jian Du, Yang Xu
Abstract
Lithology identification from well logs remains challenging due to overlapping signals and complex spatial variability in heterogeneous formations. To address this issue, a novel method, termed LIsr (Lithology Identification with a sliding window and recurrent neural network (RNN)), is proposed. The sliding window enhances distinguishing features between lithologies and reduces prediction uncertainty by incorporating label sequences predicted by an improved RNN. This network captures spatial correlations in depth between neighbor samples, a factor often overlooked in conventional machine learning (ML) approaches. A bidirectional scanning mechanism is introduced to mitigate the impact of sedimentary sequences, while residual connections and a deep architecture enhance feature extraction. The method is validated using data from the Hangjinqi Gas Field in Ordos Basin, China. Results show that incorporating unidirectional scanning improves accuracy by 5.5%, and the residual deep structure contributes an additional 22% gain. The proposed LIsr achieves 90.8% validation accuracy and 89.9% accuracy on blind wells, demonstrating its effectiveness in complex lithology prediction.