The subseasonal forecasts from the ECMWF (European Centre for Medium-Range Weather Forecasts) were used to construct weekly mean wind speed forecasts for the spatially aggregated area in Finland. Reforecasts for the winters (November, December and January) of 2016–2017 and 2017–2018 were analysed. The ERA-Interim reanalysis was used as observations and climatological forecasts. We evaluated two types of forecasts, the deterministic forecasts and the probabilistic forecasts. Non-homogeneous Gaussian regression was used to bias-adjust both types of forecasts. The forecasts proved to be skilful until the third week, but the longest skilful lead time depends on the reference data sets and the verification scores used.

Wind speed forecasts have many potential users that could benefit from skilful forecasts in different time scales, ranging from hourly to monthly forecasts. For example, short- and medium-range forecasts of extreme wind speeds are often utilised in early warnings for severe weather

Forecasts for the subseasonal time frame have improved greatly in recent years

This study concentrates on subseasonal wind forecasts in winter, as forecasts for winter in northern Europe are known to be more skilful than forecasts for other seasons

It is well known that, especially with longer lead times, the ensemble forecasts often have a systematic bias and the spread of ensemble members can be too small

This work was a part of the Climate services supporting public activities and safety (CLIPS, 2016–2018) project

The forecasts used in this study were
extended-range forecasts of 10 m wind speed,
provided by the ensemble prediction system (EPS) from the ECMWF (European Centre for Medium-Range Weather Forecasts)

Following

The same data cannot be used to both fit and evaluate the performance of the NGR, so we split the reforecasts and ERA-Interim data into two data sets: the training data set of winters starting on odd years and the validation data set of winters starting on even years. The training data set was used to fit the NGR, while the validation data set was used to evaluate the adjusted reforecasts. As the ERA-Interim data set included 31 years, the reference forecasts from ERA-Interim were based on 30 years, omitting the year under study.

We evaluated both the deterministic forecasts and probabilistic forecasts. The forecasts were the weekly means of the wind speed, spatially aggregated for the area shown in Fig.

The area used in the spatially averaged forecasts in Finland.

The effect of seasonality on the forecast skill was removed by subtracting the first three harmonics of the annual cycle.

In this study, NGR was used to correct the mean weekly forecast, not the forecasts at each time step,
as in, e.g.,

NGR provides the Gaussian probability distribution

In this study, deterministic forecasts were the mean

For the verification terminology, we follow

The continuous ranked probability score (CRPS) is used for the probabilistic forecasts

For a score

Now we can define skilful forecasts as forecasts with a skill score higher than zero.
And to be more precise, the forecast is skilful at a statistically significant level, if zero is not within confidence intervals (CIs).
As we used both Monday and Thursday forecasts, there is autocorrelation in the data, so the effective number of forecasts is not as high as 1000 for each lead time. This must be taken into account when CIs are calculated. Therefore, the CIs of verification measures are calculated with block-bootstrap

The size of the reforecast ensemble (11 members) is smaller than the size of the climatological ensemble (30 members), and the CRPS values of NGR reforecasts are not readily comparable with the climatological forecasts. Therefore, we used the formula given by

The quality of probabilistic forecasts was also evaluated using the relative operating characteristic (ROC) curves

The NGR is not a black box and the change of the constants of Eqs. (

The fitted coefficients for NGR as the function of lead time:

For the data set here,

The ME of both NGR adjusted reforecasts and climatological forecasts is nearly zero (Fig.

The verification measures of reforecasts (using the validation set, even winters 1996–2016) for the averaged area (Fig.

The correlation of NGR reforecasts decreases as the lead time increases but remains positive at the last lead time calculated (Fig.

The MSESS remains statistically significantly positive until the lead time of 21 d, when the CI includes zero (Fig.

The verification measures of reforecasts (using the validation set, even winters 1996–2016) for the averaged area (Fig.

The AUROC (Fig.

Our results are comparable to those of

Prior to analysis, we anticipated that the CRPSS (Fig.

The use of Eq. (

It is also important to further investigate the impact of the seasonal cycle on the verification results, as an uncritical reading of figures might suggest unrealistic trust in the forecasts.

The forecasts might be skilful even for the third week, but the skill is still very low, even if the skill scores are non-zero or positive. For example, an MSESS of around 0.1 can be interpreted as 10 % of the variance explained, which is very little for most applications. So it is not straightforward to see who is the potential user that could benefit from the third-week forecasts.
Using the categorisation of users by

In general, the utility of forecasts is defined by the users, so close co-operation and co-development of forecasts with the users is useful, if not essential. Moreover, the mean weekly wind itself might not be useful for most end users. For example, warnings of extreme wind would need percentiles higherthan 50 %

It seems reasonable to assume that different reanalyses generate somewhat different climatologies and observations, implying somewhat different skill scores based on different reanalyses. This is especially relevant for a variable such as wind, which is not so straight-forward to measure. So, the use of more than one reanalysis might be useful in future studies.
In this study, we used the ERA-Interim as our reference, but more recent reanalyses, such as MERRA-2

The bias-adjustment methods used here are only rudimentary and could be improved. For example,

We evaluated the weekly mean wind forecasts for Finland based on the ECMWF forecasts. The NGR was used to correct the reforecasts. The skill of forecasts appears to be positive for the third week, but the longest skilful lead time depends on the reference data sets, the scores used, and the correction methods. Also, two winters would have been a rather short time span to make meaningful inferences on how operational wind forecasts perform, so reforecasts with longer time span are essential for comparison. Even then some uncertainty remains. The needs and the competence of the end users determine whether the forecasts are useful or not. The forecasts would be most beneficial for users applying the probabilities in the decision theory framework.

The ECMWF reforecasts and ERA-Interim are available for the national meteorological services of ECMWF member and co-operating states and holders of suitable licences.

OH did most of the analysis and wrote most of the article. TKL wrote parts of the Introduction and contributed to the preparation of the paper. OR contributed to the statistical analysis. NK and VA contributed to the preparation of the paper. GH supervised the project.

The author declares that there is no conflict of interest.

Publisher’s note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

This article is part of the special issue “Applied Meteorology and Climatology Proceedings 2020: contributions in the pandemic year”.

We thank Jussi Ylhäisi for the helpful comments on the early draft of this paper.

This research has been supported by the Academy of Finland (grant nos. 303951 and 321890).

This paper was edited by Andrea Montani and reviewed by two anonymous referees.