Hybrid Deep Learning-Based Networks for Deterministic and Probabilistic Wind Nowcasting Within a Wake Vortex Warning System for Vertiport Operations

Abstract

Accurate wind prediction is essential for forecasting the transport and decay of wake vortices that might cause an adverse impact to light aircraft operations. This paper proposes an innovative framework for deterministic and probabilistic wind nowcasting, designed to be integrated into a Wake Vortex Warning System for Vertiports (WVWS-V) near major airports. The study presents and tests different deterministic nowcasting models. The superior approach employs a hybrid concept combining complex Continuous Wavelet Transform (CWT), two-dimensional Convolutional Neural Networks (2D-CNN), Long Short-Term Memory (LSTM) networks, and Light Gradient Boosting Machine (LGBM), enabling it to capture both the general patterns and abrupt fluctuations of wind occurring intermittently and repeatedly. The proposed hybrid model achieved a Mean Absolute Error (MAE) of 0.75 m/s for wind speed and 9.1° for wind direction over a 20-min prediction horizon, outperforming all other evaluated models including the statistical persistence prediction method serving as baseline. The probabilistic nowcasting establishing 95% confidence intervals of wind speed and direction utilizes a multivariate-Mixture Density Network (MDN) with weighted Mahalanobis distance. The MDN-based model provides a narrower probability range than the statistical persistence baseline model, with average improvements by 0.91 m/s for wind speed and 12.05° for wind direction. The comparison of the statistical persistence baseline and other machine learning architectures demonstrates that the proposed models achieve superior performance, enabling the WVWS-V to ensure safe and reliable light aircraft operations at vertiports.