TY - JOUR
T1 - Wind turbine load estimation using machine learning and transfer learning
AU - Xu, G.
AU - Yu, W.
AU - Kim, Taeseong
PY - 2022
Y1 - 2022
N2 - Machine learning method has always been popular to solve wind turbine related problems at a data level. However, with the limitation of the availability of relevant data, transfer learning has gained increasing attention. In this study, traditional machine learning method of artificial neural networks (ANN), together with parameter-based transfer learning method has been used to estimate wind turbine load. First, ANN load model was built for DTU 10MW wind turbine as well as NREL 5MW wind turbine. Then, parameter-based transfer learning has been applied to the above-mentioned models to estimate load for a different turbine type or two mixed turbine types. Results indicate that ANN method provides good estimation on wind turbine fatigue load. For DTU 10MW ANN model, the trend of accuracy becomes steady as the number of input samples increases and 1500 samples is deemed as the optimal number of samples for training DTU 10MW. In addition, with transfer learning, it was succeeded in building NREL 5MW model with corresponding DTU 10MW pretrained model but failed in establishing mixed dataset model neither with DTU 10MW nor with NREL 5MW pretrained model.
AB - Machine learning method has always been popular to solve wind turbine related problems at a data level. However, with the limitation of the availability of relevant data, transfer learning has gained increasing attention. In this study, traditional machine learning method of artificial neural networks (ANN), together with parameter-based transfer learning method has been used to estimate wind turbine load. First, ANN load model was built for DTU 10MW wind turbine as well as NREL 5MW wind turbine. Then, parameter-based transfer learning has been applied to the above-mentioned models to estimate load for a different turbine type or two mixed turbine types. Results indicate that ANN method provides good estimation on wind turbine fatigue load. For DTU 10MW ANN model, the trend of accuracy becomes steady as the number of input samples increases and 1500 samples is deemed as the optimal number of samples for training DTU 10MW. In addition, with transfer learning, it was succeeded in building NREL 5MW model with corresponding DTU 10MW pretrained model but failed in establishing mixed dataset model neither with DTU 10MW nor with NREL 5MW pretrained model.
UR - http://www.scopus.com/inward/record.url?scp=85131807111&partnerID=8YFLogxK
U2 - 10.1088/1742-6596/2265/3/032108
DO - 10.1088/1742-6596/2265/3/032108
M3 - Conference article
SN - 1742-6588
VL - 2265
JO - Journal of Physics: Conference Series
JF - Journal of Physics: Conference Series
IS - 3
M1 - 032108
ER -