What question did this study set out to answer?

The study aims to accurately predict the performance of large offshore wind turbines by integrating uncertainty quantification and sensitivity analysis.

June 12, 2026Open Access

Development in Surrogate-Based Polynomial Chaos with Adaptive Sobol Sensitivity Analysis for Uncertainty Quantification and Offshore 15 MW Wind Turbine Performance Prediction: Comparative, Icing, and Wind Farm Optimization Studies

Key Points

The study aims to accurately predict the performance of large offshore wind turbines by integrating uncertainty quantification and sensitivity analysis.
Developed a surrogate-based uncertainty quantification framework using a Blade Element Momentum solver and polynomial chaos expansion.
Used Sobol variance-based global sensitivity analysis on five random input parameters.
Conducted three parallel engineering studies comparing different turbine models and assessing ice accretion and wake optimization.
Achieved a maximum coefficient of performance of 0.480 and annual energy production of 71,261 MWh/year.
Identified wind speed as the primary driver of performance variance, with a Sobol index of 0.412.
Demonstrated a 18.2% reduction in performance due to severe icing and a farm efficiency of 89.6% with optimal spacing.

Abstract

Accurate performance prediction for large offshore wind turbines requires a principledtreatment of uncertainty in both the wind resource and the rotor design parameters. In thepresent work, we develop a surrogate-based, multi-level uncertainty quantification (UQ)framework coupling a physics-based Blade Element Momentum (BEM) solver with a spectralPolynomial Chaos Expansion (PCE) surrogate that replaces the expensive Monte Carloloop and apply it to the IEA 15 MW offshore reference wind turbine. The framework iscompleted by Sobol variance-based global sensitivity analysis. The contribution is methodologicalrather than algorithmic: although each individual ingredient (PCE, Sobol, BEM,and Jensen) is well established, their joint deployment in a single, internally consistent,end-to-end probabilistic workflow that simultaneously delivers (i) aerodynamic–structuralUQ with analytical Sobol ranking, (ii) a like-for-like cross-comparison of three referenceturbines, (iii) a quantitative leading-edge icing degradation study, and (iv) a farm-levelwake-steering optimization on the same IEA 15 MW reference rotor yields a unified probabilisticenvelope from which manufacturing tolerances, cold-climate investment thresholds,and farm-layout/control trade-offs can be read off consistently. Five input parametersare treated as random variables: hub-height wind speed (Weibull, k = 2.2, c = 9.8 m/s),air density, blade chord length, twist angle, and rotor speed. A degree-4 sparse PCE isbuilt by non-intrusive spectral projection using N = 5000 Sobol quasi-random realizations,which allows the Sobol indices to be recovered analytically from the expansioncoefficients at essentially no extra cost. Three parallel engineering studies complementthe core UQ analysis: (A) a head-to-head comparison of the NREL 5 MW, DTU 10 MW,and IEA 15 MW reference turbines; (B) a quantitative assessment of leading-edge ice accretionat four severity levels; and (C) a Jensen-based wake optimization for a 25-turbineoffshore array with static wake steering. The main results are as follows: the turbine reachesCp,max = 0.480 at λopt = 8.51, and an annual energy production (AEP) of 71,261 MWh/year(PCE: 70,840 ± 2,140 MWh/year, 95% CI). Wind speed emerges as the dominant driverof Cp variance (S1 = 0.412), followed by blade twist (0.198) and chord (0.143). Severeicing (30 kg/m) reduces Cp by 18.2% and increases the blade-root Damage EquivalentLoad (DEL) by 18.5%. For the array, the optimal spacing (sx = 8D, sy = 6D) gives a farmefficiency of 89.6% and 1296 GWh/year, and a 15° wake-steering offset adds a further+3.2% to farm AEP. Compared with plain Monte Carlo, the sparse PCE delivers the samestatistics with about 36% fewer model evaluations and a relative error below 0.8%.

Read Full Paperexternally

Ask AI

Helpful

Bookmark

View Full Paper