Author(s): Vincent Tjuatja; Alireza Keramat; Mostafa Rahmanshahi And Huan-Feng Duan

Linked Author(s): Huan-Feng Duan

Keywords: Physics-informed neural network Reservoir-pipe-valve-system Frequency domain Water hammer modelling

Abstract: In recent years, a new branch of surrogate modelling, the physics-informed neural network (PINN), has emerged due to the increasing availability of data and improved computing capability. Despite the growing trend of PINN modelling, the application of PINN for water hammer analysis is still relatively limited, with most applications performed in the time domain. The current study was built on the PINN approach for water hammer modelling in the frequency domain. A complex-valued neural network (CVNN) with seven hidden layers, each comprising 50 neurons, was adopted in the current study. The loss function was formulated by considering the training data and frequency-domain wave equation's physics residuals. The weighting factor for each loss term was determined based on the self-balancing algorithm. The physics-informed complex-valued neural network (PICVNN) was compared to classical CVNN, and including physics improved the model’s generalization capability. PICVNN was also exposed to three cases of epistemic uncertainties: parameter uncertainty, model uncertainty, and physics uncertainty with an unknown leak. Regardless of the inaccurate model knowledge, incorporating data can enhance the prediction result of the model.

DOI:

Year: 2025