Abstract:

The deep tight bedrock buried hills in the Central Uplift of the Liaohe Depression host fractured hydrocarbon reservoirs with significant resource potential. However, challenges arise from great burial depths, diverse lithologies, complex nonlinear relationships between fractures and logging parameters, and strong multi-solution ambiguity in fracture identification via conventional logging—resulting in low prediction accuracy. To address these challenges, the modified Long Short-Term Memory (LSTM) neural network algorithm is developed for fracture identification in deep buried hill formations. In the new algorithm, the Dropout layer is inserted between dual LSTM layers and regularization is used to mitigate overfitting; the Dense layer and the Softmax function used for classification in LSTM are replaced by Least Squares Support Vector Machines (LSSVM) with Gaussian kernel functions. Classification prediction based on features extracted by the LSTM layers can be directly approached. The proposed algorithm preserves LSTM’s sequential learning advantage for logging curves while significantly improving classification efficiency and accuracy. It effectively prevents loss of fracture-feature information and overfitting on limited training samples, while accelerating algorithm convergence. Validation results demonstrate a test-set accuracy of 91.56%, outperforming both Support Vector Machine (SVM) and standard LSTM models. This provides an efficient methodology for fracture identification in deep, complex-lithology bedrock buried hill reservoirs.

Key words: deep layer, base rock buried hill, Improved Long Short-Term Memory (LSTM) neural network, fracture identification