The stability of wellbores plays a decisive role in maintaining the safety and efficiency of drilling operations. Conventional forecasting techniques are typically dependent on empirical formulas or historical datasets, but their predictive capability is often inadequate in reservoirs characterized by strong heterogeneity and intricate geological structures. To overcome these shortcomings, this study introduces a novel approach that integrates one-dimensional convolutional neural networks (1D-CNN), bidirectional long short-term memory (BiLSTM) models, and an attention mechanism to achieve highly accurate predictions of essential petrophysical properties in undrilled formations. The parameters predicted include shale volume (Vsh), density-derived porosity (POR), compensated neutron porosity (CNCF), and the velocity ratio of compressional to shear waves (vpVsRatio). By leveraging CNN for extracting localized features, BiLSTM for capturing temporal dependencies, and the attention mechanism for highlighting critical features within logging curves, the method substantially enhances prediction performance in geologically complex reservoirs. Validation conducted on five wells within the NB19-6 block demonstrates that, relative to conventional Kriging and linear interpolation, the proposed technique reduces average prediction errors by 40%–59%, with pronounced improvements observed in deeper intervals where geological conditions are particularly challenging. This work not only supplies new methodologies for risk mitigation in drilling practices but also provides theoretical and technical foundations for precise reservoir exploitation and safe well construction in complex geological environments.