Traditional pore-pressure prediction relies heavily on empirical formulas and subjective judgment, whereas existing machine learning models exhibit limited generalization across regions and overpressure mechanisms. This study proposes an intelligent prediction method integrating feature optimization with cascaded domain-adaptive transfer learning to improve the generalizability of pore-pressure prediction. A regional dataset was first constructed from well-logging data of 14 exploratory wells in an offshore area. The continuous pore-pressure profiles used as regression targets were derived using conventional pore-pressure evaluation methods and independently assessed against available formation-pressure measurements, rather than being treated as direct pressure measurements at every depth. Hydrostatic pressure was introduced as a vertical physical constraint, and a deviation feature based on an independently constructed, target-well-excluded regional reference velocity curve was developed to reduce reliance on manually defined normal compaction trends, thereby improving inter-well feature comparability and physical interpretability. To mitigate physical inconsistencies and data distribution shifts caused by differences in overpressure mechanisms, including hydrocarbon-generation-related unloading and undercompaction, a cascaded domain-adaptation model combining canonical correlation analysis and transfer component analysis was established. This model adopts a strategy of semantic alignment followed by distribution adaptation, thereby reducing the maximum mean discrepancy between the source and target domains. Validation was conducted using blocked-depth validation, same-mechanism whole-well validation, and multiple cross-mechanism source–target transfer experiments, together with repeated perturbation-based uncertainty analysis. Results show that the proposed features and transfer-learning model improved prediction performance and physical interpretability under the tested conditions. The proposed method demonstrates potential for cross-well pore-pressure prediction under the tested multi-mechanism conditions within the study area.