Geographia Technica, Vol 17, Issue 1, 2022, pp. 92-103


Muhammad RYAN , Adhi Harmoko SAPUTRO , Ardhasena SOPAHELUWAKAN

DOI: 10.21163/GT_2022.171.07

ABSTRACT: Wind shear is one of the dangerous meteorological phenomena for aviation. This phenomenon is significant, especially at the lower level. The duration of wind shear events varies greatly, ranging from short to long. The best way to avoid accidents caused by wind shear is by predicting the event and the duration. Recent studies use Machine Learning (ML) as a nonlinear geostatistical method to predict wind shear utilizing wind observing instruments data. The data is conditioned into temporal data which is fed to the ML model. However, the ML model used is not a temporal ML model for time-series data but a generic model for a common type of data. Many studies claimed temporal models are better than generic ones to tackle temporal data. In this study, we propose Temporal Convolutional Network (TCN) to predict incoming wind shear duration and occurrence using an anemometer sensor network i.e., Low-level Wind Shear Alert System (LLWAS). The wind shear occurrence is derived from wind shear duration prediction. The proposed model is compared with other temporal models, i.e., Long-Short Term Memory (LSTM) and Gated Recurrent Unit (GRU). Different schemes of total predictor were tested to find the best predictor scheme for wind shear prediction. To measure the performance of all models in all schemes, accuracy, False Alarm Ratio (FAR), Probability of Detection (POD), and Root Mean Squared Error (RMSE) metrics are used. The result is TCN dominating almost in all metrics used i.e., Accuracy, FAR, and RMSE for all schemes against LSTM and GRU. Scheme with 4 predictors proved to bring the best performance of all models for wind shear duration prediction. The result proves TCN is the best temporal model for wind shear forecasting using LLWAS. For better wind shear duration prediction, the best scheme choice is the 4-predictor scheme.

Keywords: Wind shear, Aviation, Machine learning, Geostatistical, Temporal Convolutional Network

Full article here