Extreme precipitation is one of the most important natural hazards worldwide, both in terms of economic cost and number of people affected. Belgium is no exception, as portrayed by the July 2021 extreme precipitation event which killed 41 people, and caused over 2 billion euros worth of incurred damage in Belgium alone. As shown in a World Weather Attribution study to which RMIB contributed, such events are expected to become more likely and more intense in a warmer climate. Therefore, improving the prediction of high-impact precipitation events is a high priority of the meteorological scientific community.
Recent advances in numerical weather prediction (NWP) have greatly improved the description of precipitation. At short time scales, however, NWP is consistently outperformed by so-called nowcasting algorithms, which generally extrapolate rainfall observed by meteorological radars and other sources at high spatial and temporal resolution. Information about the predictability of the current situation is extremely valuable to improve these nowcasting algorithms, especially for state-of-the-art probabilistic or ensemble nowcasting systems.
The aim of this proposal is to improve our understanding of the short-term predictability of precipitation using novel techniques from dynamical systems theory and deep learning. Applying these techniques, inspired by new theoretical insights, will allow us to valorize the long-term archive of radar data of the RMIB in a new and innovative way to determine the dynamical properties of the observed precipitation. These insights will be used to improve the seamless prediction system of the RMIB, which aims to provide probabilistic forecasts from minutes to days ahead. Both the stochastic perturbations of the nowcasting ensemble and the blending approach to combine nowcasts and NWP can benefit from these insights.
The methodology that will be used in this research is based on very recent developments in the field of dynamical systems. This technique aims to obtain information about a system's dynamics by computing the instantaneous properties of the system's attractor. The attractor is the geometrical structure that is defined by all the states that a system can reach. Of particular interest are the instantaneous attractor dimension d, which is linked to the number of degrees of freedom necessary to describe the dynamics of the system, and the persistence θ which characterizes how long the system typically remains near that state, its predictability. The technique leverages the link between extreme value theory and the theory of recurrences to accurately compute both the instantaneous dimension and persistence along the attractor of a real system.
This technique requires long time series of measurements of a system. For any given instantaneous state of the system - or equivalently, for any given point on the attractor of the system - the extreme value distribution can be determined from the recurrences in the neighbourhood of this state. The instantaneous dimension and persistence directly follow from the fitted parameters of an extreme value distribution. This approach was able to compute very high averaged dimensions that are not accessible with traditional techniques like the correlation dimension. Moreover, it allows one to compute the instantaneous or local attractor dimension, which is much more informative than the average dimension since the dynamics of a system is known to be highly state-dependent. We aim to investigate the variability and predictability properties of precipitation by applying this technique to weather radar data.
Weather radars are recognized as the most appropriate instruments to capture precipitation at high spatial and temporal resolutions. Two different weather radar datasets will be used in this study: (1) the Belgian radar composite and (2) the European Eumetnet OPERA radar composite. Two representative radar images from this dataset are shown in the figures below.
The scientific impact of this project lies in the improved understanding of the predictability of precipitation using novel methods from dynamical systems theory. It will valorize the long climatological datasets available at the RMIB and within EUMETNET, in order to improve the understanding and forecasting of precipitation. A better knowledge of the predictability of precipitation in the nowcasting range will allow us to improve our probabilistic nowcasting system. First, it can be used to improve the stochastic perturbations in the pySTEPS system, and thus the uncertainty estimation of the precipitation forecasts. Secondly, it can be used to develop a novel blending approach between radar nowcast and NWP, in order to optimally employ all sources of information. This will be used to improve the RMIB’s seamless prediction system, which provides probabilistic forecasts from minutes to days ahead. Moreover, these developments will directly benefit the general public, who receives the forecasts through the RMIB website (200.000 users per day, with peaks up to 1 million users) and the RMIB app (over 600.000 regular users).
Dr. Martin Bonte, postdoc at RMIB
Dr. Stéphane Vannitsem, supervisor at RMIB