Data-driven optimization of Darrieus wind turbine pairs under particle-laden atmospheric conditions using LES and machine learning

This study develops a hybrid machine-learning framework that couples CFD and artificial neural networks to quantify and optimize the performance of paired Darrieus turbines in clean-air and particle-laden environments. Targeting arid, semi-arid, and coastal sites where airborne sand accelerates blade degradation, the framework resolves the influence of sand-induced surface roughness on turbine aerodynamics and power output. Clean-air CFD results are used to determine baseline optimal design parameters, which are subsequently tested under sand-particle conditions (0.5–2 mm) using LES to evaluate aerodynamic losses and erosion. A Single-Objective Genetic Algorithm (SOGA) identifies the turbine configuration that maximizes the average power coefficient (CP). Results show that larger sand particles significantly reduce performance, with a 4.8% decrease in CP at a tip speed ratio of 2.24. Regression analysis reveals a strong correlation between particle size and blade erosion rate. The results show that leading-edge erosion reduces the pressure difference across the blade, slightly increases suction-side pressure at θ = 0°, weakens the suction peak at θ = 75°, and decreases flow attachment at θ = 150°. Lift is significantly lowered in the upwind region, while delayed flow reattachment and weaker lift recovery occur in the downwind region. Cumulative erosion analysis indicates that blades exposed to 0.5 mm particles exceed the 2 mm erosion limit in under two years, while exposure to 2 mm particles leads to critical wear within months, far below the 15-year design life.