Enhancing Low-level Wind Shear Alert System (LLWAS) to Predict Low-level Wind Shear (LLWS) Phenomenon Using Temporal Convolutional Network

Abstract

Low-level Wind Shear (LLWS) is a significant phenomenon in aviation that has the potential to cause aircraft accident. To avoid the potential accident, information about the potential of LLWS occurrence ahead was needed so the pilot can avoid the area where LLWS can happen. Previous several studies used statistical model to predict LLWS. The dataset comes from the equipment system for detecting LLWS. Most of the statistical models used are Multi-Layer Perceptron (MLP) and the dataset is taken from Lidar Doppler. The approach that is often used is to transform wind data from Lidar Doppler into time series data and feed it to the MLP. For this study, the statistical model used is Temporal Convolutional Network (TCN). TCN is a dedicated time-series model. The dataset for the TCN is come from Low-level Wind Shear Alert System (LLWAS). We use the model to predict LLWS occurrence 5 minutes ahead. The feature input of TCN are wind direction and speed from LLWAS that already is being transformed and arranged to timeseries data west - east component (U) and south - north component (V). The label dataset is LLWAS's warning of LLWS occurrence data. As a comparison of the proposed model, a logistic regression model and Multi-Layer Perceptron (MLP) were also used. We also use varying lengths of input data to see how they perform against the model. The results show that TCN can outperform other comparison models with perfect recall and precision values (1) when using predictor time-series data longer than 5 minutes. This result means that the proposed model works well in predicting LLWS events using LLWAS data.