Subject

According to a report from the European Environment Agency [1], extreme weather events caused economic losses of around half a trillion in Europe between 1980 and 2020. This includes damage caused by floods, storms, heatwaves, and droughts. These estimates only capture the immediate economic damage and do not account for the human toll and loss of life resulting from extreme weather events, or delayed impacts of these events.

To improve resilience and preparedness for such extreme events, it is crucial to have timely and accurate forecasts to issue reliable early warnings. Today, machine learning based methods are developed and deployed to improve weather forecasting. These are, however, trained on the bulk of the distribution, and less on extreme events. It is therefore important to implement strategies to balance the dataset so that enough extreme events are available for training, and secondly, that the data is validated for high-impact, extreme weather.

To reach these goals, it is crucial to have a sufficiently long a historic record of extreme events. The selection of these events is not straightforward, as extremes can manifest themselves not just as storms or extreme rainfall, but also as non-events: a long drought that causes water scarcity, or a dunkelflaute (period without wind or sunshine) that causes black-outs due to a low availability of renewable energy.

In this Master’s thesis, you will define the criteria and implement methods to select extreme events from a historical and simulated future climate database. In a second part of the thesis, the link between the occurrence of these extremes and specific weather regimes [2,3] and other large-scale variables will be investigated.

The developed extreme event catalogue will be an invaluable resource: it can help train and validate machine learning methods in weather forecasting, and it can also serve as a validation set for research on numerical weather prediction models. For example, a new weather model version will be tested on the extreme event catalogue before becoming operational at the Royal Meteorological Institute.

Kind of work

Objectives

• Literature study and definition of a set of criteria that defines “extremes” of different types
  
• Programming an extreme weather detection algorithm and building the extreme event catalog based on ERA5 reanalysis data over Europe or Belgium
  
• Analysis of the occurrence of extremes related to weather regimes, using classical statistical and machine learning methods
  

Methods

• Traditional statistical methods based on extreme value theory
  
• ML-based supervised and unsupervised anomaly detection methods (e.g. based on autoencoders)
  
• Linking extreme events to weather patterns using ML
  

Framework of the Thesis

[1] EEA - Economic losses and fatalities from weather- and climate-related events in Europe. Briefing no. 21/2021. doi: 10.2800/530599. Retrieved from https://www.eea.europa.eu/highlights/economic-losses-from-weather-and on 12-05-2023.
br
[2] Neal, R., Fereday, D., Crocker, R., & Comer, R. E. (2016). A flexible approach to defining weather patterns and their application in weather forecasting over Europe. Meteorological Applications, 23(3), 389-400.
br
[3] Mastrantonas, N., Magnusson, L., Pappenberger, F., & Matschullat, J. (2022). What do large-scale patterns teach us about extreme precipitation over the Mediterranean at medium-and extended-range forecasts?. Quarterly Journal of the Royal Meteorological Society, 148(743), 875-890.

Expected Student Profile

Programming background (Python)
br
Background in machine learning
br
Experience with machine learning frameworks (e.g., TensorFlow or PyTorch)
br
Strong interest in weather, climate or hydrology