Machine Learning to improve and understand sub-seasonal forecasts of European temperature

Joachim Wilhelmus van Straaten

Research output: PhD ThesisPhD-Thesis - Research and graduation internal

654 Downloads (Pure)

Abstract

Europe faces droughts, heatwaves, and other extreme weather events every year. To prevent the worst of the impacts good probabilistic forecasts are needed, ideally issued several weeks in advance. This thesis explores whether physics-based, statistical and hybrid models can be used. Generally, sub-seasonal forecasts are difficult to make due to the chaotic evolution of the atmosphere. This thesis however shows that the division of warm and cold air over Europe can be predictable, as long as predictions are made for extended periods of time and large areas. The sources responsible for such predictability often lie in complex interactions between ocean, land and atmosphere. When disentangling these interactions with explainable Machine Learning, we learn that processes in the Pacific can be important for European summer weather. Subsequently, we evaluate ECMWF ensemble forecasts and show that large correctable errors relate to an erroneous simulation of this Pacific-to-Europe teleconnection. A correction of the physics-based forecasts with statistical post-processing increases sub-seasonal forecast skill for Europe in summer.
Original languageEnglish
QualificationPhD
Awarding Institution
  • Vrije Universiteit Amsterdam
Supervisors/Advisors
  • van den Hurk, Bart, Supervisor
  • Coumou, Dim, Supervisor
  • Schmeits, Maurice, Co-supervisor
  • Whan, K., Co-supervisor, -
Award date20 Nov 2023
DOIs
Publication statusPublished - 20 Nov 2023

Keywords

  • machine learning
  • numerical weather prediction
  • sub-seasonal
  • explainable artificial intelligence
  • heatwaves
  • Europe
  • statistical post-processing
  • atmospheric teleconnections

Fingerprint

Dive into the research topics of 'Machine Learning to improve and understand sub-seasonal forecasts of European temperature'. Together they form a unique fingerprint.

Cite this