Innovative data driven approaches to predicting long-term European wind storm risk

About the project

About the Project Background European wind storms cause insured losses of around €1.4 billion per year. However, these losses fluctuate over time, making it challenging for insurers to estimate their financial exposure. Previous work has found that year-to-year variations in European wind storms are related to patterns of large-scale atmospheric circulation. However, relatively little is known about multi-decadal variability in European wind storms and the ability of climate forecasts to predict these hazards and provide actionable information to insurers. In particular, climate models struggle to capture atmospheric circulation variability. This means estimated current and future wind storm risks based on climate models need to be carefully interpreted using a combination of physical science and decision theory. We will support you to develop an exciting project related to European wind storms that addresses cutting-edge societally-relevant research questions at the interface of climate and decision sciences. PhD Project This project is linked to a new partnership between the University of Leeds and Lloyd’s Banking Group and directly addresses real world problems. Some example topics that could be explored are: • What drives multi-decadal variability in European wind storms and can this be predicted? Understanding long timescale variability in European wind storm risk can help insurers to price policies. We can use long-term observation based datasets to characterise extreme wind storms on timescales of decades and link this with population data to estimate losses. This could focus on the statistics of wind storms or the risk of specific events analogous to observed cases. We can use uncertainty quantification methods for estimating the signal-to-noise ratio in forecasts and post-processing techniques such as calibration and ensemble matching to boost the predictable signal and make long-range forecasts more useable. • Can machine learning help to predict localised wind storms? Climate models have coarse resolution while wind storms can exhibit small scale features that contain the most damaging winds (e.g. sting jets). We could explore whether generative machine learning methods can be used to downscale climate model output to produce higher resolution forecasts. We can focus on methods such as probabilistic diffusion models and conditional generative adversarial networks that allow uncertainty information in the downscaled predictions to be quantified and where possible minimised through model training. • Can decadal climate forecasts of European wind storms support planning and decision making by insurers? All forecasts contain uncertainties which tend to grow for longer-term predictions due to weaker predictable signals. Understanding the associated uncertainties and the tolerance of users to incorporate uncertainties in their decision making is crucial to move towards climate services. Applicant Profile We are seeking students with a quantitative background (e.g. Atmospheric Science, Physics, Maths, Engineering, Computer Science) and an interest in decision science who want to apply their skills to a problem of strong societal consequence and with a direct industrial application.

Research areas