Convective-gust nowcasting based on radar reflectivity and a deep learning algorithm

Abstract

Abstract. Convective wind gusts (CGs) are usually related to thunderstorms, and they may cause great structural damage and serious hazards, such as train derailment, service interruption, and building collapse. Due to the small-scale and nonstationary nature of CGs, reliable CG nowcasting with high spatial and temporal resolutions has remained unattainable. In this study, a novel nowcasting model based on deep learning – namely, CGsNet – is developed for 0–2 h lead times of quantitative CG nowcasting, achieving minute–kilometer-level forecasts. CGsNet is a physics-constrained model established by training on large corpora of average surface wind speed (ASWS) and radar observations; it can produce realistic and spatiotemporally consistent ASWS predictions in CG events. By combining the gust factor (1.77, the ratio of the observed peak wind gust speed (PWGS) to the ASWS) with the ASWS predictions, the PWGS forecasts are estimated with a spatial resolution of 0.01∘ × 0.01∘ and a 6 min temporal resolution. CGsNet is shown to be effective, and it has an essential advantage in learning the spatiotemporal features of CGs. In addition, quantitative evaluation experiments indicate that CGsNet exhibits higher generalization performance for CGs than the traditional nowcasting method based on numerical weather prediction models. CG-nowcasting technology can be applied to provide real-time quantitative CG forecasts.

RADAR Echo Recognition of Squall Line Based on Deep Learning

“…Therefore, in recent years the meteorological community has introduced deep learning methods to achieve highly reliable and automated strong convective weather monitoring and warning such as combining deep learning with numerical weather prediction to forecast different types of convective weather [15]. For example, convective wind gust (CG) identification by CGsNet [16], and RADAR echo extrapolation by 3D-Unet-LSTM use RADAR images as training data [17]. However, all the above image-based recognition methods will lose many details after the RADAR-base data are converted into grayscale images as model input, and not enough label data are collected during the training processes.…”