Enhancing Wind Farm Efficiency through Multi-Fidelity Bayesian Optimisation

Enhancing wind farm power output is crucial for accelerating the transition to renewable energy. Turbines operating in the wake of others experience reduced wind speeds and greater turbulence, lowering power output and increasing fatigue. Wake steering, where the turbine yaw angles are adjusted to redirect wakes, can mitigate these effects and boost total farm power.

Traditionally, optimal yaw configurations are found using analytical wake models due to their computational efficiency, though they may miss important physical phenomena. Higher-fidelity large eddy simulations (LES) better capture non-linear fluid dynamics, identifying more accurate optima but at a considerably higher computational cost.

This work employs a multi-fidelity Bayesian optimisation (MF-BO) strategy, combining cheap analytical wake models (using FLORIS) with detailed LES (via XCompact3d). The MF-BO uses a non-linear autoregressive Gaussian process surrogate model to link these fidelities and an acquisition function to guide experiments, optimising learning about the optimal solution across fidelities.

This approach enables significant power improvements of the wind farm whilst conducting fewer costly LES evaluations, achieving accurate optimisation results at a reduced computational expense.