Transient simulation of multi-product pipeline driven by flow mechanisms and operational data

Objective Multi-product pipelines often operate under frequently switched conditions, making it crucial to accurately monitor flow parameters during transitions and understand the changes in hydraulic states at high and low points along these pipelines. Most existing transient estimation methods rely on precise and reliable physical models, which involve high computational costs to address multi-condition and multi-parameter combinations. In contrast, approaches based on machine learning tend to lack reliability and accuracy, as they often overlook the physical patterns associated with pipeline transients. Methods This paper presents the development of a Physics-Informed Neural Network (PINN) model for the transient simulation of multi-product pipelines driven by flow mechanisms and operational data. First, a Deep Neural Network (DNN) model was constructed to establish mapping relationships among flows, pressures, and time-space coordinates of pipeline operation. This enables the effective extraction of nonlinear correlations between flow parameters and time-space coordinates during transients. Next, the evolution of flow parameters in transients was analyzed to explore the inherent relationships, as well as the transient control equations governing these evolutions, along with the corresponding initial and boundary conditions. Finally, the transient control equations and penalty terms associated with the initial and boundary conditions were formulated using deep learning automatic differentiation, constraining model solutions within the solution space that corresponds to the transient mechanisms, thus enhancing the accuracy of the transient simulation. Results The proposed model was verified based on a simulated pipeline system under start-up, offtake, and capacity increase/decrease conditions. Compared with the DNN model, the PINN model produced predictions for G1 pipeline pressures, with the mean absolute percentage error (MAPE) reduced by 77.4%, 88.7%, and 87.8%, respectively. For flow prediction results, the MAPE decreased by 86.7%, 94.4%, and 95.7%, respectively. Subsequent verification was performed using a multi-product pipeline in Southern China under capacity increase/decrease conditions. The PINN model also outperformed the DNN model, yielding prediction results for pipeline pressures with the MAPE decreasing by 94.2% and 92.8%, respectively. Conclusion The established PINN model facilitates efficient and accurate solutions for determining flow parameters during transients under various combinations of conditions and parameters, providing support to ensure the stability and safety of pipeline operation.