Seasonal forecasting techniques were developed initially for weather extremes in tropical and equatorial latitudes, where predictability at lead times of a few months is widely recognized as having greater potential skill than in temperate latitudes. Forecasting models have been developed for extremes as spatially diverse as Atlantic hurricanes, the Indian Monsoon and Sahelian rainfall. Increasingly, however, as financial losses due to weather extremes escalate, there is interest from end-users, for example in the forestry and insurance industries, in the development of seasonal forecasting models for weather extremes in temperate latitudes.

This paper presents results from a pilot project to explore the potential for seasonal forecasting of wind storm in Europe. The predictand variables were derived from the six-hourly surface wind speeds available from the NCEP reanalyses. For each gridpoint, the 90th, 95th and 99th percentile values were calculated from daily maximum wind speed data over the period 1958 to present. Then, time series of the number of exceedances of these percentile thresholds in each winter season were created, and form the predictands. Comparison of reanalysis and observed wind speeds suggests that the reanalysis exceedance time series capture observed patterns.

Potential predictor variables are derived from two principal sources. First, deseasonalized GISST sea surface temperatures for the Atlantic Ocean and the eastern Pacific were subjected to a Principal Components Analysis (PCA), and the rotated empirical orthogonal functions form a series of statistically independent SST indices that can be used as predictors. Second, predictors were formed from relevant climate indices representing the state of the large-scale atmospheric circulation, such as the Arctic Oscillation and the East Atlantic Index. Lead times of up to a year were considered, in monthly increments, allowing for identification of the optimal lead time for each predictor variable.

In order to construct the seasonal forecasting models for wind speed exceedances in each European grid square, performances of three methods of linear regression were compared. These are: standard multiple linear (ML) regression, principal component regression and partial least squares (PLS) regression. Cross-validation (successively excluding one season for validation from the calibration) was performed. The results indicate that PC and PLS regression are superior to ML regression methods, with substantially lower standard errors of the estimates, and that PLS regression is a particularly valuable technique since it affords considerably more flexibility in experimental design than other regression methods.

Useful skill in forecasting with lead times of up to six-months was found for certain European areas. As might be expected, skill proved highest in the western maritime regions of Europe, deteriorating into central Europe. Overall, it is clear that potential exists to develop useful lead-time seasonal forecasts of wind speed extremes over western Europe for end-users in affected industries such as insurance.