T. Spangehl et al.: Intercomparing the quality of recent reanalyses for offshore wind farm planning 115 6?35?15.5?? E (https://www.fino1.de/en/location.html, last ac- cess: 6 November 2023). 3.1.1 Interannual to multi-annual variability The wind speed at heights near 100 m shows considerable in- terannual variability. During the period from 2004 to 2017 the observed annual mean wind speed at FINO1 ranges from a maximum of 10.29 m s?1 in 2007 to a minimum of 8.56 m s?1 in 2016 (Fig. 1a). The mast corrected data is not available for January 2018 to August 2019. Therefore, ob- servations 2018 onwards are not shown. The commission- ing of the wind farms Alpha Ventus in the direct vicinity east of FINO1 in 2010, Borkum Riffgrund I (2015) to the southeast and Trianel Windpark Borkum (2015) to the north- west led to a reduction of the wind speed at FINO1 (Podein et al., 2022). For this reason, the evaluation of the reanal- yses and other model-based products using FINO1 in-situ measurements as reference focuses on the undisturbed pe- riod from 2004 to 2009. For this period the data availabil- ity of the mast corrected wind speed measurements is 95 % or higher for every individual year. Here the observed an- nual mean wind speed ranges from 9.65 m s?1 in 2004 to 10.29 m s?1 in 2007 (Fig. 1b). All model-based products capture the interannual variability during the undisturbed period. Contrary to the observations all model-based prod- ucts including the global and regional reanalyses, the re- gional downscaling simulation with COSMO-CLM and the wind atlas product show maximum wind speed in 2008 and only second highest wind speed in 2007. Biases of in- dividual years range from ?0.47 m s?1 (2007) for NEWA to 0.44 m s?1 (2008) for COSMO-REA2 (Table 2). ERA5 shows a negative bias of ?0.24 m s?1 for 2004–2009. The bias is consistently negative over the individual years. A neg- ative bias of ?0.10 m s?1 is found for HoKliSim-De for the same time period. Again, the bias is consistently negative but smaller than for ERA5 for every individual year (Table 2). On the other hand, COSMO-REA6 shows a positive bias of 0.13 m s?1 which is predominantly positive for the individual years. The smallest bias of only 0.01 m s?1 for 2004–2009 is found for CERRA. Likewise, the bias is smallest for most of the individual years. For the years 2004 to 2009 there is no evidence for any systematic change of the bias over time for any of the products (Table 2). With regard to the analysis of the undisturbed time period it is concluded that the ob- servations are within the uncertainty range of the reanalyses (COSMO-REA6, CERRA, ERA5). Moreover, the regional downscaling simulation HoKliSim-De shows a smaller bias compared to the driving ERA5 dataset. Besides interannual variability the time series indicate multi-annual variations of the wind speed. Higher than av- erage wind speed values are seen for years 1998 to 2000 which are consistently represented by regional reanaly- ses (COSMO-REA6, CERRA), HoKliSim-De and ERA5 (Fig. 1a). A larger offset between ERA5 and the higher res- olution products CERRA and HoKliSim-De is seen for 1992 to 1995. Likewise, the offsets between the different products are not necessarily stable over time. For example, COSMO- REA6 shows a positive offset for most years, except for 1995, 1996 and 2017, 2018 where COSMO-REA6 shows lower wind speed values than other products. 3.1.2 Frequency distribution of wind speed and wind direction The frequency distribution of wind speed for 2004–2009 from FINO1 measurements at 102 m shows maximum val- ues betwen 8 and 11 m s?1, as indicated by the grey his- togram in Fig. 2. The regional reanalyses COSMO-REA6 and CERRA, the global reanalysis ERA5 and the regional downscaling simulation HoKliSim-De (model-based prod- ucts shown in blue) resemble the observed frequency dis- tribution of wind speed (observations shown in grey) for FINO1. For better readability of the following discussion the statistical parameters of the pdf’s shown in Fig. 2 are sum- marised in Table 3. The distribution obtained from COSMO- REA6 shows a slight shift towards higher values compared to the observations, which is manifested by a marginally higher mean value of 10.05 m s?1 compared to the observed valued of 9.93 m s?1, likewise a marginally higher median (9.75/9.6 m s?1), 99th percentile value (22.63/22.02 m s?1) and scale parameter (11.35/11.21 m s?1). COSMO-REA6 and the observations show a nearly perfect match for the shape parameter (2.25/2.26), which confirms the good agreement seen for the other parameters (Fig. 2a). CERRA shows a very good agreement with the observed mean value (9.95/9.93 m s?1) and a marginally larger median value (9.7/9.6 m s?1). The underestimation of the 99th percentile value seen for CERRA (21.75/22.02 m s?1) is slightly more pronounced than for COSMO-REA6. The shape parameter (2.3/2.26) and the scale parameter (11.23/11.21 m s?1) of CERRA are in good agreement with observations for the time period considered here (Fig. 2b). The frequency dis- tribution of ERA5 is slightly shifted towards lower values compared to observations. This shift is reflected by a slightly lower mean (9.69/9.93 m s?1), median (9.43/9.6 m s?1), 99th percentile value (21.09/22.02 m s?1), and scale parame- ter (10.94/11.21 m s?1) (Fig. 2c). The frequency distribution of HoKliSim-De is characterised by a slight underestimation of the mean (9.83/9.93 m s?1), median (9.51/9.6 m s?1), and scale parameter (11.10/11.21 m s?1). The underestimation of these parameters is less pronounced than for ERA5. The 99th percentile value (22.14/22.02 m s?1) is in good agree- ment with observations indicating that the higher resolu- tion improves the representation of extreme wind speeds (Fig. 2d). All model-based products show a slight underes- timation of the 1st percentile value compared to the observa- tions. Wind roses consisting of 16 pre-defined sectors (N, NNE, NE, ENE, E, ESE, SE, SSE, S, SSW, SW, WSW, W, WNW, https://doi.org/10.5194/asr-20-109-2023 Adv. Sci. Res., 20, 109–128, 2023