Surface Wind Speed Estimation on Multisites Anemometer Using Temporal Convolutional Network

Abstract

Surface winds in various locations are measured simultaneously using a multisite anemometer network. This network is susceptible to system failures due to sensor damage, causing a data gap during sensor removal and reinstallation. This research develops a wind speed estimation model on a multisite anemometer using the Temporal Convolutional Network (TCN) algorithm. TCN processes time domain signals in parallel, thus significantly cutting the computation time. Minutely wind speed data set was obtained from four anemometers at Juanda International Airport in Surabaya from January 1, 2022 – December 24, 2023. The model design comprises data pre-processing, dominant wind direction analysis, hyperparameter determination, training, and testing on actual data. TCN estimation models are divided into easterly, westerly, transitional, and all-directional models. These wind speed estimation models strongly correlate with actual data, with correlation coefficients of 0.70, 0.77, and 0.87. Overall, the accuracy of the TCN-based estimation model conforms to World Meteorological Organization (WMO) requirements for wind speed measurements. It achieves RMSE<5 m/s and MAE<3 m/s. As for computation duration, TCN processes the training for 87 seconds per epoch and completes the estimation in 37 seconds, much faster than CNN-BiDLSTM’'s training duration of 2206 seconds per epoch and estimation completion of 548 seconds.