Distributed learning for wind farm optimization with Gaussian processes

This article investigates optimization of wind farms using a modifier adaptation scheme based on Gaussian processes. In this scheme measurements are used to identify plant-model mismatch using Gaussian process regression, which are then used to find the optimal plant control inputs. However, for systems with many agents and a large control input space, the identification of the input-output map of the plant is challenging. Therefore, the paper proposes a distributed learning approach, in which sub-parts of the plant are identified with individual GP regression models. Afterwards, all of these are used to build a model of the overall plant-model mismatch, which is then used in the optimization. In the wind farm case the sub-parts are the individual turbines. The distributed learning approach clearly outperforms the original central learning approach in numerical illustrations of wind farm test cases.