@sebastianlerch
Professor at the Department of Mathematics and Computer Science at the University of Marburg, interested in probabilistic forecasting, statistics, ML, with applications in weather, energy, environmental sciences, and beyond
If you are going to attend EGU 2026 and are interested in statistical ML-based forecasting / nowcasting / post-processing in the geosciences, consider sumitting an abstract to our session NP 5.1, www.egu26.eu/session/55791
⏰ Deadline: January 15, 2026
The original data is available at dx.doi.org/10.35097/EOv.... We also provide code for matching forecasts and observations, and for an exemplary comparions of ML-based post-processing models.
Forecast are available for 55 meteorological variables mapped to station locations and spatially aggregated forecasts from surrounding grid points, for NWP models initialized at 00 and 12 UTC, in hourly lead times up to 21h. Observations of 6 variables are available at 170 stations.
In a new preprint available at arxiv.org/abs/2508.03845, we present the CIENS dataset, which contains contains more than 10 years of ensemble forecasts and observations from the convection-permitting ICON/COSMO system of the German meteorlogical service.
All details are available in the preprint at arxiv.org/abs/2506.03744. Python code is available at github.com/tobiasbieger....
The potential CRPS of the HRES forecast aligns well with the CRPS of the operational IFS ensemble.
AIWP models show skillful forecasts for lead times of up to 10 days when compared to the ERA5 climatology in terms of the potential CRPS.
Results on WeatherBench 2 data confirm fast-paced progress, with AIWP models, in particular GraphCast, showing improvements in the potential CRPS over the HRES model
Step 2: We then compute the CRPS on the test dataset. The resulting "potential CRPS" quantifies potential probabilistic predictive performance and serves as a proxy for the mean CRPS of real-time, operational
probabilistic products.
We propose a new measure for fair and meaningful comparisons of deterministic AIWP and NWP models:
Step 1: We subject the deterministic backbone of AIWP and NWP models post hoc to the same
postprocessing technique (isotonic distributional regression) on the test dataset.
There has been fast-paced progress in AI-based weather prediction. However, fair comparisons to physics-based NWP models are challenging:
- AI models are trained on the MSE, and might have an advantage in MSE-based comparison
- Comparisons may use different ground truth data (ERA5 vs IFS analysis)
📣 New preprint "Probabilistic measures afford fair comparisons of AIWP and NWP model output" with Tilmann Gneiting, Tobias Biegert, Kristof Kraus, Eva Walz and Alexander Jordan available at arxiv.org/abs/2506.03744. Some details below 🧵
In addition to forecast evaluation via proper scoring rules, we also evaluate the forecasts from an economic perspective by considering trading strategies that utilize the multivariate probabilistic information.
We propose a generative ML model for multivariate, probabilistic forecasting of time series of electricity prices, and compare to state-of-the-art statistical benchmark models.
New preprint: "Probabilistic intraday electricity price forecasting using generative machine learning
" available at arxiv.org/abs/2506.00044, led by Jieyu Chen
Thanks!
Thanks!
Image: https://en.wikivoyage.org/wiki/File:Marburg_Oberstadt_von_SO.jpg
Personal update: After almost 10 years at KIT, I will move to the University of Marburg as a professor at the Department of Mathematics and Computer Science in April. I will of course miss the many great colleagues and students at KIT, but am very much looking forward to exciting new opportunities.
The AI Weather Quest is now open for participation! This @ecmwf.int led competition, endorsed by the @wmo-global.bsky.social, challenges participants to push the boundaries of sub-seasonal to seasonal forecasting using artificial intelligence (AI) and machine learning (ML).
🔗 bit.ly/4b1ulaB
New preprint: "Learning low-dimensional representations of ensemble forecast fields using autoencoder-based methods" with Jieyu Chen and Kevin Höhlein: arxiv.org/abs/2502.04409. We propose dimensionality reduction methods tailored to ensemble simulations of gridded fields.
All details and links to all datasets are available in the paper (rdcu.be/d462L). Code is available at github.com/HoratN/pp-mo....
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.
Applying post-processing to the PV power predictions obtained as the output of the model chain is the most important contributor to improving the forecasts, whereas the effects of post-processing the weather inputs are negligible.
In a case study on a benchmark dataset from the Jacumba solar plant in the US, we find that post-processing generally improves the GHI and PV power forecasts. Neural network-based methods achieve slightly better performance than statistical approaches.
We investigate the use of post-processing and ML in model chain approaches, where different strategies are possible: Post-processing only the weather inputs, post-processing only the PV power predictions, or applying post-processing in both steps (or none at all).
Probabilistic PV power forecasts are often based on model chain approaches, where conversion models estimate PV generation based on weather predictions. However, weather prediction models make systematic errors and require post-processing.
Our new paper "Improving Model Chain Approaches for Probabilistic Solar Energy Forecasting through Post-processing and Machine Learning" (with Nina Horat and Sina Klerings) just appeared in Advances in Atmospheric Sciences, rdcu.be/d462L 🌤️🔌
In a case study on a benchmark dataset from the Jacumba solar plant in the US, we find that post-processing generally improves the GHI and PV power forecasts. Neural network-based methods achieve slightly better performance than statistical approaches.
We investigate the use of post-processing and ML in model chain approaches, where different strategies are possible: Post-processing only the weather inputs, post-processing only the PV power predictions, or applying post-processing in both steps (or none at all).
Probabilistic PV power forecasts are often based on model chain approaches, where conversion models estimate PV generation based on weather predictions. However, weather prediction models make systematic errors and require post-processing.