Objective The flow and pressure in the refined oil pipeline undergo rapid changes during operating condition transitions. Accurate estimation of these changes is crucial for safe pipeline operation. Current mechanism models rely on precise parameters, complicating the trade-off between solution accuracy and computational efficiency, while data-driven methods often neglect transient flow mechanisms, resulting in potential physical distortions.

Methods A hybrid modeling method integrating mechanism and data-driven approaches was proposed. Initially, non-linear dimensionality reduction was performed on high-dimensional data from pipeline stations using T-distributed Stochastic Neighbor Embedding (T-SNE), enabling the classification of quasi-steady-state and transient operating conditions. For quasi-steady-state conditions, a state estimation model was developed to calculate hydraulic parameters, complemented by a Back Propagation (BP) neural network to achieve adaptive compensation for simulation errors. Under transient conditions, mechanism-guided feature engineering was employed to explore correlations among flow parameters. A parallel Bidirectional Long Short-Term Memory (BiLSTM) network was constructed to capture time dependencies, incorporating flow time-delay characteristics. Finally, the Attention Mechanism (AM) quantified the influence weights of features on flow rate and pressure, enhancing the accuracy and interpretability of transient simulations.

Results Using a refined oil pipeline in south China as a case study, the Mean Absolute Percentage Error (MAPE) for the simulated upstream outlet and downstream inlet flow rates of Pipeline Segment A without unloading and with unloading decreased by 93.88％ and 93.61％, respectively, under quasi-steady-state conditions. During pump start-up, the MAPE for the predicted flow rates decreased by 52.06％ and 39.11％ for Pipeline Segment A, and 40.82％ and 30.87％ for Pipeline Segment B. During pump stopping, the MAPE decreased by 53.89％ and 42.62％ for Pipeline Segment A, and 59.18％ and 29.95％ for Pipeline Segment B.

Conclusion The T-SNE method defines clear boundary conditions for the model’s adaptive transitions between operating conditions. The quasi-steady-state error compensation model minimizes the frequent parameter adjustments required by traditional mechanism models while addressing the lack of physical meaning in data-driven models. By integrating BiLSTM with the Attention Mechanism (AM), feature weights are quantified, enhancing the model’s ability to characterize transient flow mechanisms. This approach accurately estimates flow rate and pressure across various operating conditions, providing a robust physical foundation for safety decision-making during transitions of operating conditions and ensuring pipeline safety.