Deep Learning and Universal Multifractal for Nowcasting Precipitation in Urban Geosciences

Abstract

Precipitation nowcasting (short-term forecasting ahead for 0 to 6 hours) is crucial for decision-making of weather-dependent industries in order to mitigate socio-economic impacts. Accurate and trustworthy precipitation nowcasting can serve as an early warning of massive floods, as well as a guide for water-related risk management. Although precipitation nowcasting is not a novel concept, it is challenging and complicated due to the extreme variability of precipitation. The traditional theory-driven numerical weather prediction (NWP) methods confront numerous obstacles, including an insufficient understanding of physical processes, enormous initial conditions impacts on predictions and requiring substantial computing resources. On the other hand, data-driven deep learning models establish a relationship between input and output data to predict future precipitation without taking into account the underlying physical processes. The framework of universal multifractal (UM) is also presented to describe the variability of precipitation nowcasting and compared to the radar observations. In this study, the convolutional long short-term memory (ConvLSTM) model is used to perform precipitation nowcasting over Metropolitan France. The study employs radar data collected every 5 minutes with a spatial resolution of 1km from Meteo-France. The preliminary results show that the structure of the field is reasonably forecast, as well as the somewhat moderate rain rates, but not the most intense ones. We discuss how to improve the methodology.