Dual-temporal prediction of wellbore condition and phase behavior during transient CO2 injection and leakage in saline aquifers

CO2 geological sequestration in saline aquifers shows significant potential for carbon reduction. However, predicting wellbore behavior during CO2 injection or leakage is complex due to flow and heat transfer mechanisms with dual temporal characteristics, both spatial distribution across depth and temporal evolution. In addition, current prediction approaches typically exhibit a singular focus on isolated sequence data, failing to account for the fundamental physical continuity inherent in depth and time series relationships. This study presents a hybrid deep learning framework, the Dual-Temporal Dual Attention Network, to predict wellbore transient temperature, pressure, and CO2 phase behavior during injection and leakage processes. A mathematical model incorporating phase transitions and thermal-hydraulic coupling mechanisms was established to characterize multiphase flow and heat transfer within wellbores. To address the dual temporal characteristics, the framework uses masked self-attention for depth sequences and temporal pattern attention for time series to capture local features, while Temporal Fusion Transformer extracts global dependencies. Model validation across representative injection and leakage scenarios substantiates exceptional predictive performance, with maximum relative errors rigorously maintained within 8% and 6% for pressure predictions, and 3% and 4% for temperature predictions, respectively. Result reveals that phase state is initially governed by axial pressure, later shifting to temperature dependence. Throughout leakage events, CO2 undergoes complex phase transitions which can amplify leakage rates through volumetric expansion. The framework provides theoretical support for CO2 sequestration risk assessment and safety management through efficient wellbore condition prediction across diverse scenarios.