Beyond Normal Weather
Probabilistic forecasting of long-term green-hydrogen resilience
Estimates how the risk of "energy droughts", long spells of low wind and sun, will change for green-hydrogen production by 2050. Global climate projections are downscaled with machine learning and turned into probabilities, not a single forecast.
Abstract
The problem and the idea
Green-hydrogen production depends on renewable energy, and renewable energy is intermittent. Feasibility studies usually rely on static historical weather, which ignores that the climate itself is shifting, and with it the odds of extended low-wind, low-sun periods. This project models the regional atmosphere as a stochastic dynamical system. Data assimilation and a lightweight machine-learning model fuse global climate projections with local observations, so the likelihood of supply failure can be estimated decade by decade.
Pipeline
How it flows
- Data
- Model / compute
- Processing
- Control step
- Output / guard
- 1 CMIP6 projections Global · coarse
- 2 ERA5 observations Local · reanalysis
- 3 ML observation operator Random forest downscaling
- 4 Monte Carlo ensemble Many 2050 trajectories
- 5 Energy-drought events Low wind + low solar
- 6 Bayesian attribution Posterior per decade
Method
How it works
-
1
Stochastic ensembles
The local climate is treated as predictable trends plus random noise. Monte Carlo ensembles map the full range of futures, with a focus on extreme low-energy events rather than averages.
-
2
ML downscaling
A simple, interpretable model (random forest or regression tree) acts as the observation operator, learning how coarse CMIP6 patterns map to precise local ERA5 weather.
-
3
Bayesian risk attribution
The drought probability is updated as a posterior across future decades, separating normal weather variability from a genuine climate-driven shift.
Results
What it delivers
- Data access is in place: ERA5 monthly observations from the Copernicus Climate Data Store and the CMIP6 catalogue via Pangeo, with a first regional case study.
- Downscaling model, ensemble generation and Bayesian attribution are in progress; results will be published here once evaluated.
Built with
- Python
- ERA5
- CMIP6
- xarray
- scikit-learn
- Bayesian inference
Acknowledgements
ERA5 data are provided by the Copernicus Climate Change Service; CMIP6 output by the World Climate Research Programme via Pangeo. The method builds on Björn Sprungk's lecture notes on probabilistic forecasting and data assimilation.
BibTeX
@misc{beyondnormalweather2026,
author = {Besharat, Parsa},
title = {Beyond Normal Weather: Probabilistic Forecasting of Long-Term Green Hydrogen Resilience},
year = {2026},
url = {https://github.com/parsabe/Beyond-Normal-Weather}
}