Statistical Post-Processing workshop - TU Delft 2019

13 November 2019

Researchers from KNMI, TU Delft, WUR, VU, and UU came together on Thursday 7 November at the Delft Institute of Applied Mathematics at TU Delft, hosted by Prof. dr. ir. Geurt Jongbloed, who also attended the workshop. They discussed the state of statistical post-processing research in The Netherlands. The researchers all work in close collaboration with KNMI to improve weather forecasts at a variety of scales. These collaborations between a government research organisation and universities that have different research focuses, from statistical to domain knowledge, are valuable as they allow the contribution from different perspectives to a common research question.

What is statistical post-processing? 
Weather forecasts are made with numerical weather prediction (NWP) models. These computer models estimate the initial state of the atmosphere, and then solve physical equations to forecast the future state of the atmosphere. The NWP models, that are run by KNMI and other government  or inter-governmental research organisations (like ECMWF), are able to skilfully forecast many weather situations. However these computer models contain errors that reduce forecast skill. The systematic errors can be corrected by statistical post-processing. 

Statistical post-processing is an area of research that focuses on improving weather forecasts using statistical or machine learning methods. Researchers find statistical relationships between observations and predictor variables from the weather model over the historical period, and use these relationships to improve future forecasts. 

Statistical post-processing at KNMI
Led by Dr Maurice Schmeits, researchers working with KNMI focus on improving weather forecasts at a number of time scales, from the next few hours to the seasonal scale, and for a variety of variables, including precipitation, wind speed, solar radiation, temperature, storm surge, and slippery roads.

“Our research focuses on better probabilistic forecasts of extreme weather, because these can enable people to take preventive action in order to reduce damage. Statistical and machine learning methods that post-process NWP model forecasts can help to improve these forecasts”, says Dr Maurice Schmeits.

The workshop attendees (left to right): Sem, Daan, Juan-Juan, Chiem, Dian, Jasper, Simon, Kiri, Maurice, Kate, and Geurt

Active Research Topics

  • Dr Kirien Whan from KNMI focuses on applying machine learning methods to improve the forecasts of extreme precipitation in the Netherlands for the first 2 days.
  • Dian Nur Ratri (PhD Student at WUR and KNMI) shows how including information from large-scale modes of climate variability can make seasonal precipitation forecasts for Java (Indonesia) that are more skilful than the raw ECMWF ensemble and to a lesser extent a simple bias correction method (empirical quantile mapping). 
  • Chiem van Straaten is a PhD student at KNMI and VU. He is working on improving forecasts in the “predictability” desert. He shows that statistical post-processing improves European sub-seasonal temperature forecasts and explores the optimum level of spatial and temporal aggregation to increase predictability. 
  • Sem Vijverberg is also a PhD student at VU. He does not use statistical post-processing but tries to improve sub-seasonal forecasts of high temperatures in the eastern US using observed soil moisture and North-Pacific sea surface temperature as predictors. 
  • Dr Kate Saunders (TU Delft) explores the dependence between storm surge and precipitation. She raises fundamental questions about best practices in statistical post-processing and develops an event based framework for thinking about storm surge in The Netherlands. 
  • Daan van Dijk is a MSc student from TU Delft who is working with KNMI to improve forecasts of cold morning road temperatures (4 - 10 am). These forecasts are important so RWS can decide if they need to salt the roads to prevent slipperiness.
  • Simon Veldkamp (KNMI; MSc student at UU) explores the value of deep learning for improving forecasts of wind speed in The Netherlands. He shows that convolutional neural networks (CNNs) are better than random forests (RF) at high wind speeds.
  • Jasper Velthoen (PhD student at TU Delft) improves the random forest machine learning method, and shows how skilful temperature forecasts can be made with less predictor variables. 
  • Last but not least: Dr Juan-Juan Cai, an assistant professor at TU Delft, addressed a different topic than the rest of the workshop attendants: she told about how she is using extreme value theory to estimate the joint probability of a strong main earthquake and a strong aftershock.