Extreme-value Neural Networks for Weather Forecasting

Abstract

Deep-learning models are commonly used in short-term precipitation forecasting. However, most deep-learning models are likely to produce blurry output problems. In order to get realistic and accurate results, AENN, a variant of Generative Adversarial Networks (GANs), has been developed. The AENN implements an additional temporal discriminator to achieve better performance on sequential-data prediction. In this thesis, we explore the use of AENN to do nowcasting for the Netherlands and surrounding area based on radar echo images. We add a self-attention module to extract long-term global and self dependencies better. In order to improve the model’s ability to predict high rain intensity, we also apply Generalized Pareto Distribution (GPD) to normalize the tail data. The proposed model is compared with PySTEPS, a state-of-the-art statistical nowcasting model, and the original AENN model without GPD normalization. The experimental results show that AENN outperforms PySTEPS in terms of instantaneous radar echo prediction in the extreme-rain period and high accumulation detection in four Dutch catchments. GPD normalization can enhance the model’s detection ability in heavy rain. However, all models still have limited overall ability on high accumulation detection and long-range prediction.