Intelligent identification of low-resistivity thin oil layers is crucial for improving logging interpretation accuracy in complex reservoirs. In Dongying formation of Chengbei 35 block, Chengdao oilfield, the low-resistivity and thin interbed characteristics lead to ambiguous logging responses and minimal differences between productive and non-productive layers. This paper innovatively applies the Gradient Boosting Decision Tree (GBDT) model for intelligent identification of lowresistivity thin oil layers. By integrating logging curve characteristics, lithoelectric test results, production data, and reservoir physical properties, a logging feature set of low-resistivity layers is constructed through mathematical feature extraction. Key discrimination parameters are selected via a decision tree feature selection mechanism as input for the GBDT model, establishing an intelligent identification model for low-resistivity reservoirs. Combined with lithoelectric test data, the reservoir lower limit standards are determined. Application results show that the GBDT model achieves an identification accuracy of 89.5%, approximately 30% higher than the traditional logging numerical model, significantly reducing errors caused by manual interpretation and providing an intelligent solution for efficient exploration and development of low-resistivity thin oil layers.

This study explores the evolutionary characteristics of microstructure, porosity-permeability, and sealing properties in caprocks (mudstone and sandstone) under thermal effects during underground coal gasification. Using mudstone and sandstone samples from the Xishanyao formation in the Baikouquan area of the Junggar Basin, we observed microstructural changes in rocks at different temperatures (25~1 200 ℃) through scanning electron microscopy. Combined with overburden pore penetration and breakthrough pressure data, we analyzed the thermal effects on rock pore networks and sealing capacity. The results show that: (1) Mudstone is rich in clay minerals, and its thermal response is mainly characterized by dehydration, cracking and melting of clay minerals, which leads to the expansion of mudstone microfractures and the growth of nascent fissures. The sandstone is rich in quartz and feldspar, and when heated, it is mainly characterized by expansion and mineral melting, fissure expansion and increased connectivity. (2) Overburden pore penetration experiments showed that mudstone porosity and permeability decreased rapidly in the range of 1~12 MPa and then stabilized, whereas the corresponding intervals for sandstone are 1~16 MPa and 16~36 MPa. The porosity and permeability of both types of rocks peaked at 900~1 200 ℃, and then decreased due to the clogging of mineral melts. (3) At 25~300 ℃, mudstone breakthrough pressure plummets due to dehydration of clay minerals. Sandstone moderately increased due to quartz thermal expansion. At 300~1 200 ℃, both dropped dramatically with the temperature increasing. After 600 ℃, caprocks completely loss their closure ability.

Lithofacies analysis serves as the foundation for identifying high-quality reservoirs. However, in areas devoid of well data or constrained by complex inter-well geological conditions, traditional techniques struggle to rapidly and accurately recognize lithofacies types and their spatial distribution. This paper proposes a convolutional neural network (CNN)-based lithofacies identification method integrated with continuous wavelet transform (CWT), achieving efficient lithofacies recognition through deep learning. Applied to the Karamay formation in the Zhengshacun area of the Junggar Basin, the methodology involves: classifying typical lithofacies based on core and logging characteristics, performing synthetic record-based well-to-seismic matching to align logging lithofacies with post-stack seismic data, converting the matched seismic waveforms into time-frequency spectrum maps using Morlet wavelet transform, generating a time-frequency spectrum dataset for different lithofacies, and constructing and training a CNN model for validation. Under horizon constraints, the planar distribution of various lithofacies is investigated. Results demonstrate that the Morlet-CNN model achieves high identification accuracy, with recognition rates exceeding 85% for 4 lithofacies types in blind well X2, significantly enhancing both the efficiency and accuracy of lithofacies identification.