Feature selection, ensemble learning, and artificial neural networks for short-range wind speed forecasts
In one sentence. ZiANN combines feature selection, ensemble learning and neural networks to post-process NWP output for six-hour nowcasts and 40-hour point wind forecasts in Austria, outperforming operational INCA and AROME.
Abstract
The objective of this study is to provide reliable nowcasting (up to six hours) to short-range wind speed forecasts of up to 40 hours ahead in 10 meters height for meteorological observation sites (i.e., point forecasting). The proposed method is a data-driven approach combining artificial neural networks, ensemble learning, and feature selection techniques. Particularly, we improve a pre-defined baseline setup using meteorological features, pre-classification by forecasting intervals, as well as spatial and temporal related data. This combination of methods is the so-called ZiANN (ZAMG interval artificial neural network) and it is optimized for both nowcasting and short-range forecasts. The developed method is one of the first machine learning based wind speed forecasts for the Austrian domain and Austrian observation sites. Heterogenous data sources are combined to derive training data for ZiANN. In particular, we consider (1) observations from weather stations and (2) output of one or several numerical weather prediction models. For (1), we use data from the TAWES network in Austria, while for (2), we use the AROME, ALARO, and/or ECMWF-IFS model interpolated for the observation site location. The model is validated by two test episodes and selected sites in Austria. Forecasts are compared to alternative methods: a random forest approach, the persistence, the currently operational nowcasting system INCA, the model output statistic META, and the NWP model AROME. Our results show that ZiANN outperforms alternative models, especially in the nowcasting-range. We conclude that machine learning techniques are suitable post-processing tools, which outperform classical methodologies.
Labels
- Topic
- Post-processing & calibration, Local & station forecasting
- Type
- New model or method
- Method
- Classical ML (trees, GP, linear, kernels)
- Time scale
- 0–6 h, 6 h–3 days
- Domain
- Local / station
- Data
- Station / in-situ, NWP analysis or forecasts (IFS, GFS, ...)
- Status
- Research
- Identifiers
- DOI:10.1127/metz/2020/1005 OpenAlex:W3010877918
- Added
- 2026-10-22
Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.