Pore Pressure Prediction in Tight Sandstone Reservoirs with a Hybrid BiLSTM-CNN Model

Abstract

Accurate pore pressure estimation in tight sandstone reservoirs is hampered by the limited sensitivity of conventional approaches, which rely on a single acoustic or resistivity log and assume shale undercompaction as the dominant overpressure mechanism. This study develops a hybrid BiLSTM-CNN model that integrates multi-parameter well-log responses to predict pore pressure equivalent density in the Upper Triassic XJH Formation of the northeastern Sichuan Basin, a tight sandstone succession deposited at the end of the Late Triassic. Downhole pressure measurements from five wells and their corresponding logging curves were compiled, yielding 2163 data points after outlier filtering and missing-value interpolation. Grey relational analysis (GRA) ranked six logging parameters—acoustic transit time, bulk density, invaded-zone resistivity, compressional slowness, deep lateral resistivity, and gamma ray—as the most informative model inputs. The network couples a two-layer BiLSTM with 64 hidden units per layer for bidirectional depth-sequence modeling and three parallel one-dimensional convolutional branches with dilation rates of 1, 2, and 3 for multi-scale local feature extraction. The model of test data achieved R2 = 0.968, RMSE = 0.0147 g/cm3, and MAE = 0.0116 g/cm3, outperforming four benchmark algorithms (BP, KNN, GBRT, and SVR). Predicted pressure profiles for a verification well reproduced lithologically controlled pressure contrasts between sandstone and mudstone members. Spatial pressure maps for the X2, X4, and X6 intervals, generated by interpolating predictions across multiple wells, show that overpressured zones coincide with greater burial depth and stronger top-seal integrity. These findings demonstrate that coupling depth-directional sequence learning with multi-scale convolutional feature extraction offers a reliable data-driven framework for pore pressure characterization in low-porosity clastic systems.