Reconstruction of proper numerical inlet boundary conditions for draft tube flow simulations using machine learning

This paper discusses an innovative strategy to determine appropriate mean and fluctuating inlet boundary conditions for the numerical flow simulations of a hydraulic turbine draft tube. Precise and comprehensive measurements at a draft tube inlet are hard to perform and thus rarely available. Moreover, its complex nature and strong sensitivity to imposed inlet boundary conditions render this flow particularly challenging to predict with numerical methods. The innovative strategy uses machine learning to properly reconstruct the incomplete or unknown inlet boundary conditions from available downstream flow information. Reynolds-Averaged Navier Stokes (RANS) and Large-Eddy Simulations (LES) turbulence modeling methods are used to investigate the draft tube and the numerical results are compared to experimental measurements. Reference simulations are first conducted with standard inlet boundary conditions commonly used in previous numerical works. For RANS, the innovative strategy significantly improves the mean velocity profiles and static pressure evolution inside the draft tube. For LES, its use with an artificial upstream extension allows a proper reconstruction of the upstream turbulent field and a more precise and realistic description of the draft tube head losses.