Stochastic Wind Field Learning Using Multi-Fidelity Surrogate Models for Robust Micro-Siting Optimization

Abstract

Micro-siting of wind farms has overriding importance on maximising the energy yield, reducing the amount of power wasted due to wind shear and improving the long term economic capability of wind energy systems. Nevertheless, stochastic and spatially heterogeneous nature of the wind fields especially across complicated terrains and amidst different weather conditions fundamentally undermines the optimization of the location of turbines. Though the high-fidelity computational fluid dynamics (CFD) simulations can effectively resolve the effects, and therefore the interactions of the wakes as well as the turbulence, its high cost renders it inapplicable in large-scale, iterative micro-siting optimization. Low-fidelity analytical wake models on the other hand are computationally efficient but are frequently without predictive accuracy and generalisation. In response to these shortcomings, in the paper, a stochastic frame work of learning wind field is established on multi-fidelity surrogate modelling and employed to conduct a sound wind turbine micro-siting optimization. The suggested methodology is a systematic way of merging comprehensive, low-fidel plasticity examinations with a restricted variety of high-fidelity CFD instabilities through an unbiased surrogate learning structure of both Gaussian process regression and deep neural networks. The formulation of multi-fidelity facilitates the effective transfer of knowledge between the levels of fidelity, and it also represents the intricate flow and interactions of the wakes. The quantification of uncertainty is clearly enshrined into the surrogate models to capture the variability of wind speed, direction, and turbulence, which enable the optimization process to capture the variability of stochastic wind behaviour and variability in performance. An optimisation approach based on robustness and using surrogate optimalisation is followed to maximise the projected energy output and minimise the sensitivity to undesirable wind realisation. It has been shown through numerical in-application on benchmark wind farm cases that the structure results in significant energy recovery, large scale capacity to reduce wind losses through wake oscillations and better capacity to maintain stability in comparison with the established deterministic and single-fidelity micro-siting designs. Besides, an order of magnitude lower input in terms of computational cost is achieved compared to CFD-only optimization without reducing the solution quality. These findings confirm that the suggested multi-fidelity stochastic surrogate model is a scalable, precise and computationally economical answer to next-generation wind farm micro-siting in the face of uncertainty.