Sebastian Lerch

PDF version: bwsyncandshare.kit.edu/s/HXJrexofCa5K…

Open 3-year postdoc position at KIT on ML for probabilistic hydro-meteorological forecasting. The position is part of an interdisciplinary research project with partners from meteorology, hydrology and federal flood forecasting agencies. Contact me if you are interested!

#ICML2024 ML4ESM workshop spotlight talk by Moritz Feik on Graph Neural Networks and Spatial Information Learning for Post-Processing Ensemble Weather Forecasts. Link to the paper arxiv.org/pdf/2407.11050

The paper is based on work Moritz did in his excellent BSc thesis. Moritz will present the paper at the workshop in Vienna on July 26, leap-stc.github.io/ml4esm-worksho….

In a case study on temperature forecasts from the EUPPBench dataset, we find substantial improvements over DRN, a state of the art neural network-based post-processing method.

New paper "Graph Neural Networks and Spatial Information Learning for Post-Processing Ensemble Weather Forecasts", with Moritz Feik and @JanStuehmer available at doi.org/10.5445/IR/100…, accepted at the ICML 2024 Workshop on Machine Learning for Earth System Modeling.

Code is available at github.com/HoratN/pp-mode…, all dataset are available under the links the paper. For related results on wind power prediction, see also doi.org/10.1002/we.2736

We further compare the post-processing approaches to a NN-based direct forecasting model, which predicts PV power based on the weather inputs without the intermediate conversion via the model chain, and achieves almost the same performance.

Probabilistic PV power forecasts are often based on model chain approaches, where conversion models estimate PV generation based on weather predictions. However, NWP models make systematic errors and require post-processing.

New preprint "Improving Model Chain Approaches for Probabilistic Solar Energy Forecasting through Post-processing and Machine Learning", with Nina Horat and Sina Klerings, arxiv.org/abs/2406.04424

Our #EGU24 post-processing and verification sessions continues today at 10:45 in room K2 with the oral session.⏰

If you are at #EGU24, don't miss the posters of our @EGU_NP post-processing and verification session tomorrow (Thursday), starting at 10.45 in Hall X3 (meetingorganizer.copernicus.org/EGU24/session/…). ⏰ Find me at X3.37 to talk about enhancing member-by-member post-processing with neural networks

On my way to #EGU24 in Vienna, where I will be until Friday. Tomorrow, Nina Horat will give a presentation on our work on UQ for AI weather models at 09:15 in room 0.11 in Session AS1.1.

New preprint on multivariate post-processing of weather regime forecasts on S2S time-scales available at arxiv.org/abs/2404.01895. Joint work with @cgrams_lsdp, @marisol_osman and Julian Quinting, led by our @KITMathSEE PhD student Fabian Mockert.

I have been appointed as a Fellow by the ECMWF, which of course is a great honor. I am looking forward to the collaborations on ML for improving weather forecasts on subseasonal-to-seasonal timescales.

Details and more results are available in the paper. Code is available at github.com/cbuelt/dduq.

Of course, results vary over variables and locations (here: CRPSS, using the ECMWF ensemble as a reference model).

📢New preprint on "Uncertainty quantification for data-driven weather models" available at arxiv.org/abs/2403.13458. Joint work with Christopher Bülte, Nina Horat and Julian Quinting. 🧵