Transfer learning strategy for data-driven correction of wind measurement around buildings using cup anemometers

Abstract

In recent years, high-precision sensors, such as ultrasonic anemometers (UAs), have been widely used for wind measurements. However, conventional sensors, such as cup anemometers (CAs), have not been replaced owing to their low cost and high robustness, and a data-driven correction method has improved their measurement accuracy. This study proposes a transfer learning strategy to improve the generalization performance and accuracy of the data-driven correction model of CA, thus reducing the cost of the collection of secondary training data. To validate its effectiveness, limited UA and CA samples as the source domain were collected at one site, and new CA samples as the target domain were collected at multiple sites. The proposed method includes the following three steps: 1) feature extraction based on wind statistical indicators; 2) feature-based clustering to categorize the source and target domain samples corresponding to different scenarios; and 3) building an artificial neural network model for each cluster of wind speed data. Two baselines, Kristensen's model and a conventional machine learning-based model, were trained with all the training data, whereas the model with the proposed strategy was selected and modeled using a fraction of the source domain samples based on clustering to adapt to the target domain samples. Overall, by comparing the error evaluation metrics and relative error of the wind speed statistics, it was verified that the correction accuracy of the proposed strategy was superior to the two baselines, specifically in representing the wind speed fluctuation characteristics that were not captured in the CA measurements.