Joint prediction and real-time warning of pipeline and valve-chamber pressures are essential for the safe and efficient operation of natural gas transmission systems. Traditional physical modeling has difficulty capturing the nonlinear and nonstationary characteristics of pressure dynamics, and existing data-driven approaches are often limited in modeling the complex spatiotemporal coupling between pipelines and valve chambers. To address these issues, a dynamic convolutional bidirectional long short-term memory (DCNN-BiLSTM) method was proposed for pressure prediction based on real operational data. The DCNN component adapted convolutional kernel weights to input fluctuations, improving the extraction of nonstationary and local dynamic features, while the BiLSTM component models bidirectional temporal dependencied to better represent time-series patterns. The method achieves an coefficient of determination R² of 0.998 and an root mean square error of 8.632×10^-3 on the pressure prediction task, performing favorably compared with several commonly used baseline models. Building on this, a dynamic warning mechanism was designed to adaptively update pressure thresholds using operational data, achieving an anomaly detection accuracy of 98.23%. The developed pressure prediction and early-warning system has been deployed in real operational environments, contributing to improved intelligent monitoring and safety management in natural gas transmission systems.