Abstract/Summary

We present a regime‐switching vector autoregressive method for very short‐term wind speed forecasting at multiple locations with regimes based on large‐scale meteorological phenomena. Statistical methods for wind speed forecasting based on recent observations outperform numerical weather prediction for forecast horizons up to a few hours, and the spatio‐temporal interdependency between geographically dispersed locations may be exploited to improve forecast skill. Here, we show that conditioning spatio‐temporal interdependency on “atmospheric modes” derived from gridded numerical weather data can further improve forecast performance. Atmospheric modes are based on the clustering of surface wind and sea‐level pressure fields, and the geopotential height field at the 5000‐hPa level. The data fields are extracted from the MERRA‐2 reanalysis dataset with an hourly temporal resolution over the UK; atmospheric patterns are clustered using self‐organising maps and then grouped further to optimise forecast performance. In a case study based on 6 years of measurements from 23 weather stations in the UK, a set of 3 atmospheric modes are found to be optimal for forecast performance. The skill of 1‐ to 6‐hour‐ahead forecasts is improved at all sites compared with persistence and competitive benchmarks. Across the 23 test sites, 1‐hour‐ahead root mean squared error is reduced by between 0.3% and 4.1% compared with the best performing benchmark and by an average of 1.6% over all sites; the 6‐hour‐ahead accuracy is improved by an average of 3.1%.