Region-based convolutional neural network for wind turbine wake characterization from scanning lidars

Abstract A convolutional neural network is applied to lidar scan images from three experimental campaigns identify and characterize wind turbine wakes. Initially developed as a proof-of-concept model single data set in complex terrain, the now improved generalized two other unique sets, one located near an escarpment offshore. The model, initially using scans collected predominantly westerly flow, exhibits sensitivity flow direction. thus successfully through implementing standard rotation process before input into ensure westerly. sample size of used train increased, along with generalization process, these changes are shown enhance accuracy robustness when characterizing dissipating asymmetric Applied offshore which nearly 20 wakes included per scan, 95% success rate 74% wake fragments. generalize well new although increase characterization offset by false positive identifications.