Abstract
Off-grid alpine communities face significant energy transitions due to their reliance on imported fossil fuels or unstable regional grid connections. While micro-hydro power (<100 kW) provides a promising, environmentally benign solution, standard turbine runner geometries perform poorly in ultra-low-head (<3 m) and highly variable seasonal streamflow regimes. This study presents a machine learning (ML)-optimized micro-hydro turbine blade design framework tailored for low-head run-of-river installations in decentralized alpine regions. Integrating a 3D Reynolds-Averaged Navier-Stokes (RANS) solver with a Gaussian Process surrogate model driven by Bayesian acquisition functions, we optimized a multi-parameter blade geometry to maximize hydraulic efficiency across dual operating conditions: peak spring runoff and reduced winter baseflows. Computational fluid dynamics (CFD) predictions were validated via a 1:5 scale physical model tested in a calibrated hydraulic flume, followed by a six-month field deployment in a remote alpine stream in the Val d'Hérens, Switzerland. The ML-optimized blade geometry achieved a peak hydraulic efficiency of 88.4% at a 2.2 m design head, representing a 14.2% efficiency gain over conventional NACA 4412 baseline profiles. Under off-design low-flow conditions, efficiency improvements reached 22.8%, successfully mitigating flow separation and localized cavitation. Field monitoring confirmed continuous power generation averaging 18.6 kW with robust voltage stability across seasonal flow drops. These findings demonstrate that machine learning-assisted hydrodynamic shape optimization can overcome historical low-head efficiency penalties, enabling resilient, decentralized micro-hydro infrastructure in alpine terrains.
