Pirone, Dina (2023) Probabilistic rainfall nowcasting with Machine Learning models. [Tesi di dottorato]

Anteprima

Testo Pirone_Dina_35.pdf Download (4MB) | Anteprima

Abstract

Nowcasting models use real-time data to predict rainfall with short lead times - from a few minutes up to six hours. They influence many aspects of daily life in hydrological, agricultural, and economic sectors. For example, they facilitate drivers by predicting road conditions, enhance flight safety by providing weather guidance, and prevent casualties by issuing rainfall alerts which can affect human life and cause environmental issues. However, short-term prediction is challenging because meteorological variables are strongly interconnected and rapidly change during events. In addition, the long computational times and low spatial and temporal resolution of nowcasting models do not often suit the short-term prediction requirements. This thesis focuses on developing an approach for probabilistic rainfall nowcasting with machine learning. Since machine learning does not require any previous physical assumption, this research investigates their ability to provide reliable and quick forecasts. A machine learning model for probabilistic rainfall nowcasting for short lead times - from a few minutes up to 6 hours - is proposed. The model employs cumulative rainfall fields from station data as inputs for feed-forward neural networks to predict rainfall intervals and the corresponding probabilities of occurrence. Using cumulative rainfall depths from station data overcomes the lack of temporal memory of the feed-forward neural networks. In this way, using only the current rain field as input, the model exploits pattern recognition techniques combining temporal - cumulative rainfall depth - and spatial - cumulative rainfall field – information. Several feed-forward neural networks were independently trained and tested on almost 360 rainfall events over the study area – one of the eight warning zones of the Campania Region. First, comprehensive nowcasts verifications were performed to analyze probabilistic nowcasts' reliability using continuous and categorical indicators. The performance of the models was also compared with the results of two different benchmarks: Eulerian Persistence and Pysteps. Then, to assess the extendibility of the procedure to other regions, the model was applied to another study area that differed from southern Italy one: the Flanders Region of Belgium. Results showed that using temporal and spatial information enables the model to predict short-term rainfall using only the current measurements as input, resulting in a rapid, easily replicable, and convenient nowcasting approach. Therefore, the procedure effectively predicts multi-step rainfall fields and is suitable for operational early warning systems.

Downloads

Actions (login required)

Modifica documento