Abstract
To reduce the cost of training data collection in the data-driven calibration of cup anemometer, this study proposes a novel strategy using the transfer learning based on domain adaptation to apply existing data appropriately to model training of new prediction. The proposed calibration framework performs an unsupervised step with relevant features-based clustering analysis and a supervised step with several artificial neural network models. Two field measurements were taken at two locations around a building, about a year apart, in which the former was set as the train datasets and the latter was set as the test datasets. The prediction accuracy of the models trained by the existing train datasets to the new test datasets was evaluated. The comparative study between k-means clustering and agglomerative clustering was also discussed. Overall, by comparing with the evaluation metrics on the error evaluation, and the relative error of wind speed statistics of calibrated measurements, it was verified that the calibration effect of the proposed strategy outperformed conventional machine learning method which utilized the train datasets directly.
