1. Introduction
With the vast expansion of global wind energy resources in the past two decades, wind power plant sites have increasingly penetrated regions of cold climate, including higher elevations in the midlatitudes. The total installed wind capacity in cold climates across North America, Europe, and Asia was estimated at 127 GW at the end of 2015 and is projected to increase to approximately 186 GW, or 30% of global installed capacity, by the end of 2020 (Lehtomäki 2016). The advantages of operating wind farms in cold climates, such as good resource availability and low population density, however, come at the cost of operating in harsh weather conditions.
In the wind energy context, cold climate regions are defined as areas that frequently experience atmospheric icing or temperatures below typical operational limits (Lehtomäki et al. 2018). Several types of atmospheric icing exist: icing in cloud or fog occurs when supercooled liquid cloud droplets impinge on a surface and freeze upon contact; freezing precipitation happens when raindrops fall on a surface with a temperature below the freezing point; wet snow can stick to surfaces and accrete at temperatures slightly above freezing. Ice accretion on wind turbine blades deteriorates their aerodynamic properties resulting in reduced power output. High accumulated ice loads can lead to structural damage of the turbine rotor. Last but not least, there is the risk of ice fall and ice throw for people in their surroundings (Krenn et al. 2018).
Forecasts of turbine icing potentially allow wind farm operators or energy traders to take protective measures or adapt to power production losses. A sizeable body of literature on icing prediction has formed in the last 10–15 years. Vassbø et al. (1998), Drage and Hauge (2008), Nygaard (2009), and Thompson et al. (2009) were among the first to use the output of numerical weather prediction (NWP) models to forecast icing on structures explicitly, by coupling them to an icing model developed by Makkonen, Finstad, Stallabrass, and others (Makkonen 2000, and references therein; referred to as the Makkonen model hereinafter). Liquid water content (LWC) was found to be the most crucial parameter for the accurate modeling of ice loads, requiring microphysics schemes capable of retaining LWC in the air even at temperatures well below the freezing point. Subsequent works explored increasingly higher model resolution as well as different combinations of physics parameterizations and compared model results to explicit measurements of LWC, droplet sizes, and ice load (Nygaard et al. 2011; Podolskiy et al. 2012; Yang et al. 2012; Davis et al. 2014; Thorsson et al. 2015). Of the cloud microphysical schemes tested, the scheme by Thompson et al. (2008) emerged as the most capable at predicting LWC. Almost all authors, however, noted the great sensitivity of icing forecasts to model parameter uncertainties, an aspect addressed systematically only recently (Molinder et al. 2018, 2019).
One of the goals of ice forecasting is the explicit prediction of ice-related power production losses. This involves the identification of turbine icing periods from meteorological observations or measured power curves (Arbez et al. 2016; Davis et al. 2016a,b) and the modeling of the quantitative impact of icing on power production, requiring suitable statistical models calibrated on actual wind farm data (e.g., Davis et al. 2016b; Scher and Molinder 2019; and contributions to Winterwind International Wind Energy Conference).
With the availability of new technologies for deicing and anti-icing of turbine blades (cf. Lehtomäki et al. 2018), proactive measures against turbine icing and resultant power losses or downtimes have become feasible. These solutions are valuable particularly in more densely populated regions, such as continental Europe, where regulations often require wind turbines to be shut down at the first signal of ice buildup on blades to prevent ice throw and to be restarted only after lengthy deicing procedures (cf. e.g., Bredesen et al. 2017a; Krenn et al. 2018). This makes turbine operations especially sensitive to the icing process, even if incidences of icing are generally less severe and shorter lasting than for northern regions.
The ICE CONTROL project, conducted in 2016–19 by the Zentralanstalt für Meteorologie und Geodynamik (ZAMG), the University of Vienna, the private weather service provider Meteotest, and the Austrian power supplier VERBUND, aimed at improving icing forecasts to enable optimized—possibly automated—operations of wind farms under icing conditions. In support of model development and validation, a two-winter field campaign was conducted in 2016/17 and 2017/18 at a wind farm in Germany, including standard meteorological measurements as well as camera image analyses of icing on the turbine hub and blades. In addition, measurement data from another wind farm in the Czech Republic were obtained.
This paper presents a verification study of the skill and potential economic value of deterministic forecasts of ice growth (also, active icing; Bredesen et al. 2017b). The phase of active ice accumulation on blades has been associated with the strongest production losses (e.g., Bernstein et al. 2012; Bergström et al. 2013); it is also the sensitive phase during which preventive anti-icing can make a difference. Icing forecasts for the range up to day 3 are produced from global and limited-area high-resolution models coupled to the Makkonen model. Verification serves to address the following questions: (i) What is the reference skill of coarse-resolution global operational models for forecasting icing and prerequisite conditions, such as freezing temperatures and high relative humidity? Are high-resolution limited-area models superior? (ii) Do forecasts of active ice growth have potential economic value for the decision-making by wind farm operators? (iii) Can simple, inexpensive extensions to deterministic point forecasts, such as gridpoint neighborhood ensembles, improve icing predictions?
A distinguishing feature of this study is the longer-term, two-winter verification of icing forecasts, which, to our knowledge, has been attempted in only a few other works so far (Arbez et al. 2016; Davis et al. 2016b; Molinder et al. 2019). The record length allows assessing the statistical confidence in the results. Verification results also provide a baseline against which any forecast improvements—such as through the use of ensembles, explored in a follow-up study—have to be tested.
The rest of this paper is organized as follows: In section 2, surface and upper-air observations and site-specific icing measurements used for verification are described and site climates are discussed. Section 3 details the setup of NWP and icing models. Two case studies presenting one correct and one incorrect forecast at a wind farm are discussed in section 4. Section 5 presents results of a regional-scale verification as well as of verification at icing measurement sites. Findings are summarized and conclusions are drawn in section 6.
2. Observations
a. Surface and upper-air observations
The purpose of regional-scale verification in this study is to characterize the average forecast errors of two global models, the European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecast System (IFS) and the U.S. National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS), and one limited-area model, the Weather Research and Forecasting (WRF) Model. Comparison with verification results at the wind farm sites allows to discern possible local effects on systematic errors (e.g., orographic complexity).
Figure 1b shows the locations of the 194 surface stations and 14 upper-air (radiosonde) observation sites used for regional-scale model verification. The area covered by the observations corresponds to the inner WRF modeling domain (cf. section 3b). Data were collected for the two winter periods November 2016–March 2017 and November 2017–March 2018. Surface and upper-air measurements were obtained, respectively, from the Integrated Surface Database (Smith et al. 2011, https://www.ncdc.noaa.gov/isd) of the U.S. National Centers for Environmental Information and from dataset 337.0 (NCEP/NWS/NOAA/U.S. Department of Commerce 2008) of the Research Data Archive (https://rda.ucar.edu) of the U.S. National Center for Atmospheric Research. Surface observations at hourly interval were subsampled to 3-hourly resolution to match the available global model output (cf. section 3a). Upper-air observations are available at 12-hourly intervals.
Maps showing the extents of WRF domains (a) D01 and (b) D02 as well as surface and upper-air stations used for verification. In (b), the western and eastern triangle symbols indicate, respectively, the locations of wind farms Ellern (Germany) and Kryštofovy Hamry (Czech Republic) at which icing measurements are verified.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Maps showing the extents of WRF domains (a) D01 and (b) D02 as well as surface and upper-air stations used for verification. In (b), the western and eastern triangle symbols indicate, respectively, the locations of wind farms Ellern (Germany) and Kryštofovy Hamry (Czech Republic) at which icing measurements are verified.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Maps showing the extents of WRF domains (a) D01 and (b) D02 as well as surface and upper-air stations used for verification. In (b), the western and eastern triangle symbols indicate, respectively, the locations of wind farms Ellern (Germany) and Kryštofovy Hamry (Czech Republic) at which icing measurements are verified.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
b. Icing measurement sites
A two-winter field campaign was conducted in 2016/17 and 2017/18 at wind farm Ellern (EL), located on the Hunsrück Range in Rhineland–Palatinate, Germany (Fig. 1b). The Hunsrück is a low mountain range west of the city of Mainz stretching ~100 km from southwest to northeast. Instruments were located on a hilltop on the nacelle of an Enercon E-126 turbine at 780 m MSL. Table 1 gives a summary of the instrumentation at EL. Beyond standard meteorological sensors to measure temperature, humidity, and wind, two types of icing measurements were made.
Coordinates of site locations, types of wind turbines, and names of instruments used at icing measurement sites. Winters 1 and 2 respectively refer to periods November 2016–March 2017 and November 2017–March 2018.
Three heated Mobotix AG M15 secure “webcam” systems were installed at the turbine hub, two pointing toward rotor blades, one pointing toward the instrumentation on the nacelle (Fig. 2 and Table 1). A heating and ventilation system ensured continuous operation during icing episodes. Images were recorded at 10-min intervals. A camera image analysis by Meteotest provided five classes of the severity of instrumental icing (light, light-moderate, moderate, moderate-heavy, and heavy) and 3 classes of ice growth (light, moderate, strong). For the categorical assessment of ice growth, the change of instrumental icing was observed. The image analysis was conducted manually and cross validated between several analysts for selected periods. A quantitative estimation of both ice load and ice growth would have been desirable, in particular with regard to icing forecast evaluation and calibration [section 5c(2)]. However, this was impeded by the predominantly light to moderate icing conditions observed, which would have required a finer resolution of ice load changes, and by occasional issues with image quality.
Meteotest camera images of measurement instrumentation at the hub of a turbine at wind farm EL (780 m MSL) at (a) 1220 UTC 30 Nov 2016, (b) 1200 UTC 4 Jan 2017, and (c) 1200 UTC 25 Jan 2017. The times of images (b) and (c) are indicated as vertical long-dashed lines in Figs. 5k and 5l, below.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Meteotest camera images of measurement instrumentation at the hub of a turbine at wind farm EL (780 m MSL) at (a) 1220 UTC 30 Nov 2016, (b) 1200 UTC 4 Jan 2017, and (c) 1200 UTC 25 Jan 2017. The times of images (b) and (c) are indicated as vertical long-dashed lines in Figs. 5k and 5l, below.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Meteotest camera images of measurement instrumentation at the hub of a turbine at wind farm EL (780 m MSL) at (a) 1220 UTC 30 Nov 2016, (b) 1200 UTC 4 Jan 2017, and (c) 1200 UTC 25 Jan 2017. The times of images (b) and (c) are indicated as vertical long-dashed lines in Figs. 5k and 5l, below.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
A Combitech AB IceMonitor (Combitech 2016; Meteotest 2016), a 0.5-m-long and 3-cm-wide vertical, freely rotating cylindrical rod, was used to measure ice load. The instrument, however, did not operate continuously during the two winters. Therefore, in this work, data collected by the IceMonitor serve only for case study analysis (section 4).
An additional dataset was made available through a cooperation agreement with Enercon GmbH for a wind farm located in the Ore Mountains, close to the village of Kryštofovy Hamry (KH; Fig. 1b). The Ore Mountains stretch ~130 km from southwest to northeast along the border between Czech Republic and Germany. Turbine hub heights in the KH wind farm reach up to 930 m MSL, that is, ~150 m higher than at EL. Sensor data and camera image analyses were available from the top of an Enercon E-82 wind turbine and a meteorological measurement mast (Table 1).
c. Site climates
Because of their slightly different elevations and locations in central Europe, EL and KH exhibit differences in their site climates. Figure 3 summarizes these differences in terms of a two-season climatology of four types of conditions, contiguous periods of which are referred to here as weather events: 1) freezing temperatures (T ≤ 0°C), 2) freezing and humid conditions (T ≤ 0°C and RH ≥ 85%), 3) active icing (ice growth class ≥ light), and 4) instrumental icing (class ≥ light).
Histograms of weather event duration at wind farms EL (blue lines) and KH (red lines) during periods November 2016–March 2017 and November 2017–March 2018 for weather events (a) T ≤ 0°C, (b) T ≤ 0°C and RH ≥ 85%, (c) active ice growth on turbine hubs, and (d) instrumental icing on turbine hubs (as derived from camera images). Histograms represent the event duration sums for each duration bin; numbers give the event numbers for each bin. Bins represent intervals with closed lower and open upper bounds. The sample climatological rates of occurrence (number of hours during which a particular condition was met divided by total number of hours in the period considered, excluding events shorter than 1 h) are denoted as sEL and sKH for each condition and site combination.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Histograms of weather event duration at wind farms EL (blue lines) and KH (red lines) during periods November 2016–March 2017 and November 2017–March 2018 for weather events (a) T ≤ 0°C, (b) T ≤ 0°C and RH ≥ 85%, (c) active ice growth on turbine hubs, and (d) instrumental icing on turbine hubs (as derived from camera images). Histograms represent the event duration sums for each duration bin; numbers give the event numbers for each bin. Bins represent intervals with closed lower and open upper bounds. The sample climatological rates of occurrence (number of hours during which a particular condition was met divided by total number of hours in the period considered, excluding events shorter than 1 h) are denoted as sEL and sKH for each condition and site combination.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Histograms of weather event duration at wind farms EL (blue lines) and KH (red lines) during periods November 2016–March 2017 and November 2017–March 2018 for weather events (a) T ≤ 0°C, (b) T ≤ 0°C and RH ≥ 85%, (c) active ice growth on turbine hubs, and (d) instrumental icing on turbine hubs (as derived from camera images). Histograms represent the event duration sums for each duration bin; numbers give the event numbers for each bin. Bins represent intervals with closed lower and open upper bounds. The sample climatological rates of occurrence (number of hours during which a particular condition was met divided by total number of hours in the period considered, excluding events shorter than 1 h) are denoted as sEL and sKH for each condition and site combination.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Freezing and humid conditions, a prerequisite for icing, are generally more frequent at KH (sample base rate of sKH = 44%) than at EL (sEL = 26%) (Fig. 3b). Active ice growth occurs roughly 3 times as often at KH (sKH = 15%) as at EL (sEL = 6%). Also, persistent periods of ice growth are more common at KH than at EL, with peaks in event duration in the ranges of 12–24 h at KH versus 4–8 h at EL (Fig. 3c). Differences for instrumental icing are less pronounced (Fig. 3d), since, once ice has accreted on the structure of the turbine hub, it remains until the subsequent warm period. The exception are the longest icing events (≥96–168 h), which are less frequent at the lower-elevation EL site. Table 2 permits further differentiation of the icing severities affecting the wind farms. While a range of instrumental icing severities from light to moderate was observed during the two winters, light severities of ice growth dominated the samples at both sites (>85% at EL; >66% at KH). This has implications for the stratification of verification results given the small sample sizes [cf. section 5c(2)].
Distribution of severities of instrumental icing and ice growth observed at wind farms EL and KH, averaged over periods November 2016–March 2017 and November 2017–March 2018. The last line for each wind farm gives the sums for icing categories light and greater. Percentage numbers are provided with respect to the total hours in the periods indicated.
Overall, results show that the two sites belong to ice classes 2–3 (EL) and 4 (KH), according to the IEA ice classification for wind energy sites (Bredesen et al. 2017b). In a preliminary analysis of winters 2013–15, the predominant icing type at EL was found to be in-cloud icing, responsible for >80% of ice-related turbine shutdowns (Weissinger 2017). This finding likely applies to KH as well, given the geographic proximity of the sites.
3. Numerical weather prediction models and icing model
a. Global models IFS and GFS
During the two winter seasons examined in this study, the IFS Cycle 43 and the GFS versions 13 (first winter) and 14 (second winter) were operational. Over continental Europe, the horizontal resolution is about 8 km for IFS (O1280 grid) and about 13 km for GFS (T1534 grid). The models differ by the number of vertical levels (137 for IFS; 64 for GFS). In both cases the vertical coordinate is hybrid, transitioning from terrain-following at the ground to pressure-based in the higher atmosphere. Although specific implementation details differ, the philosophy of most physical schemes is largely similar in the two models. Complete details and literature references can be found on the models’ documentation web pages (ECMWF 2020; NCEP/NOAA 2020).
Differences between microphysical schemes are particularly relevant for this study. Both IFS and GFS adopt bulk microphysical schemes, which are described in detail by Forbes et al. (2011) for IFS and by Zhao and Carr (1997) for GFS. The IFS and GFS schemes differ by the number of hydrometeor classes considered. The IFS microphysics scheme includes separate prognostic equations for the mixing ratios of cloud liquid water, cloud ice, rain, and snow. The GFS microphysics scheme only considers one predictive variable, cloud mixing ratio, which accounts for both liquid water and ice. Differences in microphysical schemes imply that the coupling between meteorological and icing models is implemented in slightly different terms for IFS and GFS (section 3c).
b. WRF Model
The Advanced Research version (version 3.9) of the Weather Research and Forecasting Model (WRF-ARW; Skamarock and Klemp 2008) is used in this study to produce high-resolution forecasts. The model setup mimics that of the 13-km Rapid Refresh (RAP) and 3-km High-Resolution Rapid Refresh (HRRR) forecasting systems, which are run operationally at NOAA/NCEP (Benjamin et al. 2016).
Table 3 summarizes the details of the model setup. The WRF Model is run in a one-way nested two-domain configuration, with horizontal grid spacings of 12.5 and 2.5 km in the outer and inner domains (Fig. 1a). Vertical model levels are finely spaced (~20 m) close to the surface and stretched toward the model top. Initial and boundary conditions are provided by IFS or GFS. The model is initiated at 0000 UTC and run for a forecast horizon of 72 h every day in the two winter periods November 2016–March 2017 and November 2017–March 2018. Hourly model output is produced.
Details of the WRF-ARW, version 3.9, model setup.
No data assimilation beyond that of the global models is used. However, soil temperature and moisture fields are recycled from one run to the next to permit a shorter spinup of near-surface moisture fields in the first forecast hours. The topography in the WRF Model is corrected at 10–20 grid points surrounding the wind farm sites to prevent a gross underestimation of the real topography, in particular in the coarse model domain. The correction procedure was carried out carefully by hand ensuring that topography gradients do not significantly exceed those produced by the WRF Preprocessing System (WPS) to avoid numerical instabilities in the model simulations.
Previous modeling studies of icing have shown the microphysics parameterization to be crucial for quantitatively accurate icing forecasts (e.g., Nygaard et al. 2011; Davis et al. 2014). Here, we adopted the “aerosol aware” scheme by Thompson and Eidhammer (2014), which extends the original scheme by Thompson et al. (2008) by the explicit prediction of number concentration of cloud droplets and cloud-condensation and ice nuclei. The adoption of a two-moment scheme for cloud water is a significant extension relative to the abovementioned previous studies, in which a prescribed droplet number concentration was found to be a limiting factor in modeling the case-to-case variability of median volume diameter, a crucial parameter in icing models.
c. Icing model
Since the efficiency factors depend on ρ, V, A, and M, the model is nonlinear and has to be solved numerically. Integration of Eq. (1) yields M for a reference cylinder (here, 3 cm in diameter and 1 m in length) at any desired time. Ice load and thickness are initialized with their values at +24 h of the previous forecast, to account for icing episodes lasting beyond the forecast horizon. Also, a simple ice-shedding component is used, similar to that by Davis et al. (2014), in which all ice is removed from the cylinder when temperature exceeds 1°C for more than 1 h.
Although all of the necessary parameters for the modeling of accretion of supercooled liquid water are predicted by WRF, Nc was not available for either global model and only the total liquid water density ρtot was available for GFS. Therefore, Nc was set to a constant 150 cm−3 when coupling the Makkonen model to IFS and GFS forecasts. For GFS, the icing model was driven with the total liquid water density ρtot.
4. Case studies
In this section, we present two case studies meant to highlight the difficulty of accurately forecasting icing events at a given site. Both cases occurred at EL and were associated with significant ice accretion on the turbine nacelle and blades (Figs. 2b,c) and turbine shutdowns. The cases have been selected to contrast a correct icing forecast (hit) with an incorrect icing forecast (miss).
Figure 4 shows forecast charts extracted from WRF-IFS D01 and D02 at an altitude of 780 m MSL. The prominent synoptic feature of the first case, 3–4 January 2017, is a low pressure system moving from southern Scandinavia over the Baltic Sea (Fig. 4a). The system brings northwesterly flow over northwestern Europe, including the area around EL. Radiosonde ascents from nearby station Idar-Oberstein, 40 km southwest of EL, reveal freezing temperatures and a saturated air mass from 1800 UTC 3 January to 1200 UTC 4 January 2017 (not shown). Forecasts agree with this observation, showing significant LWC in the region at turbine height (Fig. 4c).
Forecast charts at an altitude of 780 m MSL for case studies (left) 3–4 Jan 2017 and (right) 24–25 Jan 2017 for (a),(b) temperature (color shading), isobars (contours), and wind (barbs) extracted from WRF-IFS D01 and (c),(d) LWC (color shading, isotherms (contours), and wind (barbs) extracted from WRF-IFS D02. Gray areas mark regions with terrain that is higher than the altitude displayed. The red star indicates the location of wind farm EL in Germany.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Forecast charts at an altitude of 780 m MSL for case studies (left) 3–4 Jan 2017 and (right) 24–25 Jan 2017 for (a),(b) temperature (color shading), isobars (contours), and wind (barbs) extracted from WRF-IFS D01 and (c),(d) LWC (color shading, isotherms (contours), and wind (barbs) extracted from WRF-IFS D02. Gray areas mark regions with terrain that is higher than the altitude displayed. The red star indicates the location of wind farm EL in Germany.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Forecast charts at an altitude of 780 m MSL for case studies (left) 3–4 Jan 2017 and (right) 24–25 Jan 2017 for (a),(b) temperature (color shading), isobars (contours), and wind (barbs) extracted from WRF-IFS D01 and (c),(d) LWC (color shading, isotherms (contours), and wind (barbs) extracted from WRF-IFS D02. Gray areas mark regions with terrain that is higher than the altitude displayed. The red star indicates the location of wind farm EL in Germany.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
In contrast, the second case, 24–25 January 2017, is characterized by a stable high pressure system over large parts of the continent with weak pressure gradients (Fig. 4b). EL is affected by weak northeasterly flow; widespread stratus with low cloud base, visible in radiosonde ascents (not shown), again causes wind turbines to be immersed in clouds. This is in agreement with forecasts predicting comparatively lower but nonnegligible LWC at the site (Fig. 4d).
Figure 5 overlays measured and forecast meteorological and icing parameters for the two cases. Temperature forecasts show some deviations from the observations (Figs. 5a,b). In the first case, the forecast predicts too high temperatures starting at around 1800 UTC 3 January 2017, leading to an early end of the modeled icing event at ~0500 UTC 4 January 2017 (Fig. 5j). In the second case, the forecast temperature is up to 3°C lower than observed. Wind speed and direction are forecast relatively well in both cases (Figs. 5c,d). The cloud-base height (CBH; Figs. 5e,f) or the availability of LWC (Figs. 5g,h) are among the most critical parameters for icing. In both cases, the CBH is located at around model topographic height and below turbine hub height. For the first case, the model reproduces the magnitude of ice load on the IceMonitor (Fig. 5k). In the second, forecast ice growth rate (Fig. 5j) and ice load (Fig. 5l) are almost zero, in contrast to observations.
Observations and WRF-IFS D02 forecasts for the two case studies: (left) 3–4 Jan 2017 and (right) 24–25 Jan 2017 at wind farm EL. (a),(b) Forecast (solid or dashed lines) and measured (filled or open circles) temperature (blue) and relative humidity (red); (c),(d) forecast (solid or dashed lines) and measured (filled or open circles) wind speed (blue) and wind direction (red); (e),(f) forecast and measured CBH (blue solid lines and filled circles, respectively); (g),(h) forecast cloud LWC (blue solid line) and forecast cloud droplet number concentration Nc (red dashed line); (i),(j) forecast ice growth rate on reference cylinder (blue solid lines) and ice growth class derived from camera images (red circles); (k),(l) forecast and measured ice load on reference cylinder (blue solid lines and filled circles, respectively) and instrumental icing derived from camera images (red open circles). In (e) and (f), horizontal lines indicate model topographic height (~600 m MSL) and turbine hub height (780 m MSL). In (k) and (l), long-dashed vertical lines indicate times for which forecasts (Fig. 4) and camera images (Figs. 2b,c) are shown. Short-dashed vertical lines indicate times of icing-related turbine shutdown.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Observations and WRF-IFS D02 forecasts for the two case studies: (left) 3–4 Jan 2017 and (right) 24–25 Jan 2017 at wind farm EL. (a),(b) Forecast (solid or dashed lines) and measured (filled or open circles) temperature (blue) and relative humidity (red); (c),(d) forecast (solid or dashed lines) and measured (filled or open circles) wind speed (blue) and wind direction (red); (e),(f) forecast and measured CBH (blue solid lines and filled circles, respectively); (g),(h) forecast cloud LWC (blue solid line) and forecast cloud droplet number concentration Nc (red dashed line); (i),(j) forecast ice growth rate on reference cylinder (blue solid lines) and ice growth class derived from camera images (red circles); (k),(l) forecast and measured ice load on reference cylinder (blue solid lines and filled circles, respectively) and instrumental icing derived from camera images (red open circles). In (e) and (f), horizontal lines indicate model topographic height (~600 m MSL) and turbine hub height (780 m MSL). In (k) and (l), long-dashed vertical lines indicate times for which forecasts (Fig. 4) and camera images (Figs. 2b,c) are shown. Short-dashed vertical lines indicate times of icing-related turbine shutdown.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Observations and WRF-IFS D02 forecasts for the two case studies: (left) 3–4 Jan 2017 and (right) 24–25 Jan 2017 at wind farm EL. (a),(b) Forecast (solid or dashed lines) and measured (filled or open circles) temperature (blue) and relative humidity (red); (c),(d) forecast (solid or dashed lines) and measured (filled or open circles) wind speed (blue) and wind direction (red); (e),(f) forecast and measured CBH (blue solid lines and filled circles, respectively); (g),(h) forecast cloud LWC (blue solid line) and forecast cloud droplet number concentration Nc (red dashed line); (i),(j) forecast ice growth rate on reference cylinder (blue solid lines) and ice growth class derived from camera images (red circles); (k),(l) forecast and measured ice load on reference cylinder (blue solid lines and filled circles, respectively) and instrumental icing derived from camera images (red open circles). In (e) and (f), horizontal lines indicate model topographic height (~600 m MSL) and turbine hub height (780 m MSL). In (k) and (l), long-dashed vertical lines indicate times for which forecasts (Fig. 4) and camera images (Figs. 2b,c) are shown. Short-dashed vertical lines indicate times of icing-related turbine shutdown.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
The differences between the two cases and the related model performances at forecasting icing seem to be explained by the modeled cloud droplet concentrations Nc. In the first case, during the phase with strong ice growth (1800 UTC 3 January to 0200 UTC 4 January), Nc is relatively low, approximately 100–150 cm−3 (Fig. 5g). Given the considerable amount of available LWC, this translates into large droplets that imply a high collision efficiency α1 in the Makkonen model. In the second case, LWC reaches similar levels, even if only for a short period (1600–1800 UTC 24 January). However, high modeled values of Nc on the order of 300–400 cm−3, associated with smaller droplets with low collision efficiencies, result in negligible ice rate.
The marked differences in Nc between the two cases are physically rooted in the “aerosol aware” microphysics scheme by Thompson and Eidhammer (2014). The aerosol content of air masses depends on their origin and history. The first case is associated with strong flow from the northwest, that is, with a clean maritime air mass from the North Sea, characterized typically by low aerosol loads and low droplet concentrations. The second case exhibits weak low-level flow capped by a strong
Overall, the two cases seem to differ markedly in the microphysical properties of the cloud water causing icing. We speculate here that conditions for the second case, characterized in the model by weak wind and slow ice growth, transient peaks of LWC, and small droplet sizes, reside in a sensitive regime of the icing model in which the forecast fails to reproduce the observed ice accretion. However, to pinpoint the exact mechanisms is beyond the reach of this study, since explicit measurements of LWC and Nc were not conducted at the Ellern site and have only recently become feasible (e.g., Jokela et al. 2019; Rydblom et al. 2019; Kaikkonen et al. 2020).
Therefore, despite the adoption of explicit two-moment prediction of cloud water, icing forecasts can still drastically miss individual cases. This underlines the need for statistical evaluation over multiple events to determine their longer-term skill, which is the focus of the rest of this paper.
5. Forecast verification
In this section, we present the verification of forecasts for two winter seasons, November 2016–March 2017 and November 2017–March 2018, or a total of 302 days. Regional-scale verification at surface and upper-air stations serves to validate model setups and characterize average forecast errors of temperature T, dewpoint temperature Td, and wind speed V. The main focus, then, is on the forecasts of icing and associated meteorological conditions at wind turbine hub height.
a. Verification scores
Therefore, H quantifies the fraction of correctly forecast events with respect to the total of observed events, and F represents the fraction of observed nonevents for which an incorrect “yes” forecast was issued. PSS > 0 indicates positive model skill, where skill refers to the ability to correctly forecast occurrence and/or nonoccurrence of weather events more often than would be expected by chance (Hogan and Mason 2012). Bias B is the ratio of “yes” forecasts to observed events and is used to assess the degree of overforecasting (B > 1) or underforecasting (B < 1). H, F, PSS, and B are all standard metrics for verification of binary events and are nicely related graphically using relative operating characteristics (ROC) diagrams, with H and F spanning the diagram axes and PSS and B being drawn as isolines [cf. section 5c(1)]. GSS instead is consulted for its robustness for rare events (i.e., low s), with GSS > 0 indicating positive skill. For a detailed discussion of verification measures, their statistical properties, and caveats, refer to Wilks (2011) and Hogan and Mason (2012).
Because of finite sample size, all verification metrics represent sample statistics with inherent uncertainty, which is estimated here using blockwise bootstrapping of the verification scores to determine their 95% confidence interval (cf. Wilks 1997, 2011). A bootstrap sample size of n = 1000 and a block size length of 30 days, based on the decorrelation time scale of observed T, Td, and V time series, are used for all continuous variables and binary weather events.
b. Regional-scale verification
A marked diurnal cycle in bias is evident in the forecasts of surface parameters (2-m temperature and dewpoint temperature, 10-m wind speed) by all models (full-color lines in Figs. 6a–c). To a lesser extent, bias oscillations with diurnal periodicity are detectable even in the 925 hPa forecasts (Figs. 6d–f), at least for T and Td. The 925-hPa pressure level corresponds roughly to the hub height of turbines in the wind farms considered and is referred to as such in the following. Noticeable intermodel differences exist both in the mean bias value and in the amplitude and phase of the diurnal bias variability. The daily averaged temperature bias is generally close to 0, with the exception of the hub-height forecasts by GFS and IFS, which are slightly positively biased. Dewpoint bias is markedly positive for the WRF simulations, both at the surface and at hub height. It is instead negative for the global models, with the exception of the hub-height GFS forecasts (slightly too moist). Wind speed bias is close to zero for the global models, while it is markedly negative (positive) at the surface (hub height) for WRF.
Average forecast biases of (left) temperature, (center) dewpoint temperature, and (right) wind speed at surface and upper-air measurement stations in WRF domain 2 (Fig. 1) in the two-winter verification period, with the biases being represented as a function of forecast lead time for (a)–(c) surface level and (d)–(f) hub height (approximated by the 925-hPa pressure level) and as a function of (g)–(i) forecast values at hub height. Green lines are IFS forecasts, orange lines are GFS forecasts, blue lines are WRF forecasts coupled to IFS, and red lines are WRF forecasts coupled to GFS. The dashed and solid lines respectively represent WRF forecasts with 12.5-km (D01) and 2.5-km (D02) horizontal grid spacing. The legend in (g) applies to all panels.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Average forecast biases of (left) temperature, (center) dewpoint temperature, and (right) wind speed at surface and upper-air measurement stations in WRF domain 2 (Fig. 1) in the two-winter verification period, with the biases being represented as a function of forecast lead time for (a)–(c) surface level and (d)–(f) hub height (approximated by the 925-hPa pressure level) and as a function of (g)–(i) forecast values at hub height. Green lines are IFS forecasts, orange lines are GFS forecasts, blue lines are WRF forecasts coupled to IFS, and red lines are WRF forecasts coupled to GFS. The dashed and solid lines respectively represent WRF forecasts with 12.5-km (D01) and 2.5-km (D02) horizontal grid spacing. The legend in (g) applies to all panels.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Average forecast biases of (left) temperature, (center) dewpoint temperature, and (right) wind speed at surface and upper-air measurement stations in WRF domain 2 (Fig. 1) in the two-winter verification period, with the biases being represented as a function of forecast lead time for (a)–(c) surface level and (d)–(f) hub height (approximated by the 925-hPa pressure level) and as a function of (g)–(i) forecast values at hub height. Green lines are IFS forecasts, orange lines are GFS forecasts, blue lines are WRF forecasts coupled to IFS, and red lines are WRF forecasts coupled to GFS. The dashed and solid lines respectively represent WRF forecasts with 12.5-km (D01) and 2.5-km (D02) horizontal grid spacing. The legend in (g) applies to all panels.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Some aspects of the diurnal variability of model biases lend themselves to causal explanation. For instance, in the WRF forecasts, negative daytime peaks in T bias correlate with positive peaks in Td bias. Since no data assimilation is performed in these runs, this correlation is likely indicative of shortcomings of the surface energy balance modeling, in particular of the partitioning between sensible and latent heat fluxes at the surface. Excessive daytime evapotranspiration, at the expense of sensible heating, explains both the generally positive Td bias and the opposite V biases at the surface and hub height (weakened convective mixing implies both too weak winds near the surface and too high winds in the middle of the boundary layer).
The amplitude of bias oscillations at hub height is smaller than at the surface. This might depend on the reduced impact of imperfect surface-layer modeling at this altitude, but possibly also on the lower temporal resolution of the verifying measurements, which allows only a marginally adequate sampling of the diurnal cycles. Hub-height T biases are close to zero for the WRF runs and positive for the global models (Fig. 6d). The opposite is true for Td biases (Fig. 6e). Again, the correlation between T and Td biases can be attributed to the modeling of land–atmosphere interactions. Wind speed forecasts have a near-zero bias for the two global models and a positive bias of 1 m s−1 for WRF, both approximately constant in time (Fig. 6f).
Probing deeper into the origin of diurnally varying model biases is out of the scope of this study, so we do not speculate further on the issue. Instead, in order to understand whether such biases may impact icing modeling, we consider more carefully forecasts at hub height. Figures 6g–i shows how bias changes as a function of forecast values. The most relevant parameter ranges for wind turbine icing (especially in-cloud icing) correspond to T and Td around the freezing point and wind speed above the cut-in wind speed of 3 m s−1. In these ranges, the T bias is close to zero (slightly positive) and the Td bias is positive (close to zero) for the WRF runs (the global models). The wind speed bias is more markedly positive for WRF. These findings suggest that, without prior correction of model error, turbine icing would be more frequently observed in WRF simulations, for the following reasons. Their T bias at the freezing point is near zero, as opposed to slightly positive (thus unfavorable for ice accumulation) for the global models. Their Td bias is positive, implying more frequent saturation at near-freezing temperatures compared to the global models. Last, their higher V bias implies on average greater cloud water fluxes and ice mass accretion on turbine blades.
We finally notice that, for the WRF simulations, biases depend only moderately on model resolution and on the global driving model. Using Δx = 2.5 km and nesting into IFS generally seems to have a beneficial impact, in particular on surface-level forecasts (Figs. 6a,b), but this is not always the case. For instance, surface wind speed biases are less negative in WRF-GFS than in WRF-IFS runs (Fig. 6c), and T and Td biases at hub height seem unaffected by model resolution (Figs. 6a,b,d,e).
c. Verification of icing measurements at turbine hub height
Prior to forecast verification at icing measurement sites, measurement data were subsampled from their native temporal resolutions (1–10 min, Table 1) to hourly resolution to meet the hourly (3 hourly) output frequency of WRF (IFS and GFS). For T, Td, and V, the 10-min mean preceding each full hour was taken. For ice growth, the 1-h maximum preceding each full hour was used. The focus of on-site verification is on the forecast range 12–36 h, based on the time frame of decision-making for end users, for whom forecasts are mostly available starting only from lead time ~12 h (because of the time it takes for NWP models to be integrated, forecasts to be distributed to users, etc.). Observation-forecast pairs from that forecast range are lumped together to evaluate mean verification scores. Results for higher lead times, for example, 36–60 h, have also been considered and model skill has been found to be insignificantly degraded.
For ease of presenting the results, we will focus our attention mainly on wind farm EL in the next sections, using the results at KH for corroboration. A direct comparison between both sites is provided in section 5c(4).
1) Comparison of raw and bias-corrected forecasts
Figure 7 shows the biases of forecasts of T, Td and V across their value ranges as determined from a two-winter verification at EL for forecast lead times 12–36 h. The raw model forecasts exhibit considerable errors. Forecasts of T (Fig. 7a) tend to be too cold for all models over most of the temperature range, importantly in the range T ≤ 2°C, with the WRF D01 and D02 forecasts having the strongest bias around the freezing point. Similarly, forecasts of Td (Fig. 7b) tend to be too dry for the range Td ≤ 5°C, with the two global models showing the largest deviations. Outliers or peaks in both T and Td toward the limits of their ranges should be interpreted carefully, since they are based on only a small sample of values. Last, all models exhibit an underestimation and gross overestimation of wind speed, respectively, for low and high wind speeds (Fig. 7c).
Average biases of (a)–(c) raw forecasts and (d)–(f) bias-corrected forecasts as a function of forecast values for forecast range 12–36 h at turbine hub height for wind farm EL (780 m MSL) in the two-winter verification period, showing (left) temperature bias, (center) dewpoint temperature bias, and (right) wind speed bias.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Average biases of (a)–(c) raw forecasts and (d)–(f) bias-corrected forecasts as a function of forecast values for forecast range 12–36 h at turbine hub height for wind farm EL (780 m MSL) in the two-winter verification period, showing (left) temperature bias, (center) dewpoint temperature bias, and (right) wind speed bias.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Average biases of (a)–(c) raw forecasts and (d)–(f) bias-corrected forecasts as a function of forecast values for forecast range 12–36 h at turbine hub height for wind farm EL (780 m MSL) in the two-winter verification period, showing (left) temperature bias, (center) dewpoint temperature bias, and (right) wind speed bias.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
The comparison of biases at EL (Figs. 7a–c) with domainwide averages (Figs. 6d–f) shows largely consistent behavior. Biases in T and Td have similar trends and differ only by an offset, the EL forecasts being slightly colder and drier. This is likely influenced by the choice of altitude at which forecasts are extracted (here, 780 m MSL). Differences in wind speed biases are more marked (strong overestimation by the model, e.g., by 4 m s−1 at 15 m s−1, as compared with <2.5 m s−1 for domain averages). Two other causes for this come to mind: (i) insufficient model representation of surface roughness at the hilltop (the immediate surroundings of the turbine are characterized by coniferous woodland with ~20 m canopy), (ii) turbulence or mean-wind deficits related to turbine wakes within the wind farm, which are not explicitly modeled. However, site-specific biases are corrected statistically in the following, so we do not further explore their origin here.
The biases in T and Td are also reflected in ROC diagrams in Fig. 8. In these diagrams, diagonal dashed gray lines and dashed blue lines represent, respectively, isolines of PSS and frequency bias B. PSS isolines to the upper left of the zero-skill line (PSS = 0, dashed black) mark the region of positive model skill with respect to random forecasts. Under- and overforecasting occur, respectively, to the lower-left and upper-right of the B = 1 isoline (also see section 5a). The slightly colder bias in WRF D02 forecasts results in an overprediction of freezing conditions at the site, visible in the relatively larger false alarm rate and frequency bias B > 1 (Fig. 8a). The combination of T and Td biases leads to pronounced differences at predicting freezing and humid conditions between coarse and high resolution (Fig. 8b). Being too dry, IFS and GFS forecasts significantly underpredict these conditions, while WRF D02 tend to overpredict due to their cold bias.
ROC diagrams of (a)–(c) raw forecasts and (d)–(f) bias-corrected forecasts for forecast range 12–36 h in the two-winter verification period, showing (left) weather condition T ≤ 0°C, (center) T ≤ 0°C and RH ≥ 85%, and (right) active ice growth—all measured and forecast at turbine hub height at wind farm EL (780 m MSL). The coloring of the forecast models is as in Fig. 7. Horizontal and vertical bars indicate the 95% confidence intervals of H and F as determined from blockwise bootstrapping (section 5a); F is limited to the range of 0–0.35 for better visual separation of nearby points, in contrast to the usual display of ROC diagrams showing equal ranges 0–1.0 on both axes. Diagonal gray dashed lines represent isolines of the Peirce skill score (PSS = H − F) with values increasing in steps of 0.2 toward the upper-left corner, with the solid black line corresponding to no skill (PSS = 0). Diagonal dashed blue lines represent isolines of forecast frequency bias B, with values provided with each isoline.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
ROC diagrams of (a)–(c) raw forecasts and (d)–(f) bias-corrected forecasts for forecast range 12–36 h in the two-winter verification period, showing (left) weather condition T ≤ 0°C, (center) T ≤ 0°C and RH ≥ 85%, and (right) active ice growth—all measured and forecast at turbine hub height at wind farm EL (780 m MSL). The coloring of the forecast models is as in Fig. 7. Horizontal and vertical bars indicate the 95% confidence intervals of H and F as determined from blockwise bootstrapping (section 5a); F is limited to the range of 0–0.35 for better visual separation of nearby points, in contrast to the usual display of ROC diagrams showing equal ranges 0–1.0 on both axes. Diagonal gray dashed lines represent isolines of the Peirce skill score (PSS = H − F) with values increasing in steps of 0.2 toward the upper-left corner, with the solid black line corresponding to no skill (PSS = 0). Diagonal dashed blue lines represent isolines of forecast frequency bias B, with values provided with each isoline.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
ROC diagrams of (a)–(c) raw forecasts and (d)–(f) bias-corrected forecasts for forecast range 12–36 h in the two-winter verification period, showing (left) weather condition T ≤ 0°C, (center) T ≤ 0°C and RH ≥ 85%, and (right) active ice growth—all measured and forecast at turbine hub height at wind farm EL (780 m MSL). The coloring of the forecast models is as in Fig. 7. Horizontal and vertical bars indicate the 95% confidence intervals of H and F as determined from blockwise bootstrapping (section 5a); F is limited to the range of 0–0.35 for better visual separation of nearby points, in contrast to the usual display of ROC diagrams showing equal ranges 0–1.0 on both axes. Diagonal gray dashed lines represent isolines of the Peirce skill score (PSS = H − F) with values increasing in steps of 0.2 toward the upper-left corner, with the solid black line corresponding to no skill (PSS = 0). Diagonal dashed blue lines represent isolines of forecast frequency bias B, with values provided with each isoline.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Given the evident dependence of forecast skill on model biases, an assessment and ranking of models according to their raw performances would not be meaningful. Thus, before further analysis, we perform a binwise bias correction, that is, we subtract the respective mean error in each bin from the forecast values. The correction removes most of the systematic forecast error across the parameter ranges (Figs. 7d–f), except where it is too poorly defined due to insufficient sample size. As a consequence, forecasts of all bias-corrected models cluster on the ROC diagram for freezing and freezing and humid conditions overlapping at similar PSS and B (Figs. 8d,e). This is true, to a lesser degree, even for conditions of ice growth (Fig. 8f, discussed in more detail in the next section).
2) Skill of icing forecasts and dependence on ice growth rate threshold
Camera image analyses are used as a basis for model verification. These analyses yield three severity classes for active icing (section 2b); however, in the two-winter period considered here, light icing conditions predominated at both wind farms (cf. Table 2 and section 2c). Therefore, only nonzero icing conditions and associated model performance can be studied with reasonable accuracy, without differentiating between icing severities.
Prior to verification of icing forecasts, the threshold of modeled ice growth rate above which active icing is forecast to occur, needs to be selected, since it impacts verification results. Figure 9 shows B, PSS, and GSS of icing forecasts as a function of the icing threshold. For all models and for both sites, the scores display a strong dependence on the threshold. The lower the icing threshold, the more often icing conditions are forecast, which is reflected in B increasing monotonically with decreasing threshold and reaching significant levels of overforecasting at low thresholds (Figs. 9a,d). PSS and GSS, on the other hand, are seen to saturate (Figs. 9b,c,e,f), indicating the limit of optimizing model skill by lowering the threshold and thereby increasing the sensitivity of the icing forecast. It is evident that a unique threshold optimizing the performance of all models does not exist. This is not surprising given the large differences in horizontal resolutions and microphysical schemes and consequently differing model climatologies of LWC. To permit a fair comparison, we select model-dependent optimal thresholds consistent with verification scores. For each model, that threshold is chosen that optimizes PSS and GSS (at their plateaus) but keeps the frequency bias below a reasonable value of
Skill scores of 12–36-h icing forecasts as a function of ice growth rate threshold for wind farms (a)–(c) EL and (d)–(f) KH, showing (left) forecast frequency bias B, (center) PSS, and (right) GSS. For WRF-IFS and WRF-GFS forecasts, dashed and solid lines refer to D01 and D02 forecasts, respectively. The three vertical dashed black lines refer to ice growth rate thresholds of 5 × 10−5, 1.0 × 10−4, and 2.5 × 10−4 kg m−1 h−1.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Skill scores of 12–36-h icing forecasts as a function of ice growth rate threshold for wind farms (a)–(c) EL and (d)–(f) KH, showing (left) forecast frequency bias B, (center) PSS, and (right) GSS. For WRF-IFS and WRF-GFS forecasts, dashed and solid lines refer to D01 and D02 forecasts, respectively. The three vertical dashed black lines refer to ice growth rate thresholds of 5 × 10−5, 1.0 × 10−4, and 2.5 × 10−4 kg m−1 h−1.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Skill scores of 12–36-h icing forecasts as a function of ice growth rate threshold for wind farms (a)–(c) EL and (d)–(f) KH, showing (left) forecast frequency bias B, (center) PSS, and (right) GSS. For WRF-IFS and WRF-GFS forecasts, dashed and solid lines refer to D01 and D02 forecasts, respectively. The three vertical dashed black lines refer to ice growth rate thresholds of 5 × 10−5, 1.0 × 10−4, and 2.5 × 10−4 kg m−1 h−1.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Figure 8f shows ROC diagrams for 12–36 h forecasts of ice growth. Generally, all model forecasts exhibit relatively limited but positive skill with hit rates of 0.3–0.4, false alarm rates on the order of 0.05, PSS of 0.25–0.35, and B ≈ 1. The highest hit rates and PSS are obtained for high-resolution WRF D02 forecasts, surpassing coarser-resolution forecasts in the mean. However, the margin of improvement compared to the other model variants is small and, taking confidence intervals into account, the differences between models are hardly statistically significant.
Note that, in two related studies, verification of icing forecasts has been conducted in a fashion similar to this work. Davis et al. (2014) predicted blade ice load >0.001 kg in a one-month period at a wind farm in central Sweden and obtained much higher values of H and PSS, but also F (approximately 0.9, 0.5, and 0.4, respectively). However, accumulated ice load was predicted rather than instantaneous ice growth, which is more easily affected by forecast timing errors. Arbez et al. (2016) produced icing forecasts at three wind farms in eastern Canada during a two-winter period and yielded values of H and B in relatively wide ranges of 0.4–1.0 and 0.8–3.0, respectively, for two different models. Icing events in the models were identified through a single ice growth rate threshold of 10−2 kg m−1 h−1. In both studies, deviations from the results presented here are quite large. Apart from differences in model performance, this could be attributed to the differing definitions of icing and associated forecast thresholds but also to the background icing climatologies at all sites (icing being much more frequent and more severe at the Swedish and Canadian sites). Results, therefore, do not seem to be directly comparable; instead, they need to be evaluated in the context of each site and of the application envisioned by the authors.
3) Skill of forecasts depending on temporal resolution
The limited skill of icing forecasts (Fig. 8f) relative to temperature and dewpoint criteria (Figs. 8d,e) points to the great challenge of predicting precisely in space and time the conjunction of conditions causing ice accretion. So far, we have verified forecasts hour by hour with observations at the exact forecast validity time. However, predictions even of periods in which icing should be expected can present useful information for forecast users, guiding, for instance, their anti-icing procedures (switching on/off of blade heating systems). As an example, a relaxed timing of ±3 hours is considered, that is, a model forecast is said to be correct if a weather condition was both observed and forecast at least once in a 6-h time span.
Figure 10 shows the comparison of forecast performances for exact and relaxed timings in terms of ROC and PEV diagrams. For clarity, results are shown for IFS and WRF-IFS D01 and D02 only, those for GFS and WRF-GFS D01 and D02 differing only insignificantly. Differences in skill between the exact and relaxed event timings are almost invisible for freezing conditions (Fig. 10a), which can be attributed to the already high model skill and long-lasting freezing periods (cf. Fig. 3a), which are relatively easy to capture. Instead, freezing and humid conditions are forecast with higher H and PSS (Fig. 10b). The latter is also reflected in PEV diagrams with all model variants showing higher value at cost-loss ratios above the sample climatological rate (i.e., to the right of the vertical black dashed line, Fig. 10e). The greatest improvement is seen for conditions of ice growth (Figs. 10c,f), with a considerable increase in H (by ~0.2 or ~50%) and PSS (by ~0.15 or ~40%) for all model variants. By virtue of Eq. (6), this results in a broadening of the corresponding PEV curves and significant increase in maximum value (≡PSS) for the relaxed timing constraint. This means that economic benefit can be drawn from these forecasts in a broader range of cost-loss ratios, making them potentially more valuable to users. At the same time, the ranking of models is unaltered, with icing being increasingly better forecast from IFS to WRF-IFS D01 to WRF-IFS D02, visible, for example, at PSS values. In the rest of this work, we will show results for the relaxed timing in the verification of icing conditions but stick to the exact timing for temperature and humidity.
Comparison of performances of bias-corrected 12–36-h forecasts at wind farm EL for exact hourly timing of weather conditions and relaxed 6-hourly timing (“±3 h”), showing (a)–(c) ROC diagrams and (d)–(f) PEV diagrams. On ROC diagrams, horizontal and vertical bars indicate the 95% confidence intervals as determined through blockwise bootstrapping. In PEV diagrams, 95% confidence intervals are indicated in transparent color backgrounds to mean PEV curves. On ROC diagrams, isolines of frequency bias B = 1 are drawn as blue solid and dashed lines for exact and relaxed timings, respectively; F is limited to the range of 0–0.35 for better visual separation of nearby points.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Comparison of performances of bias-corrected 12–36-h forecasts at wind farm EL for exact hourly timing of weather conditions and relaxed 6-hourly timing (“±3 h”), showing (a)–(c) ROC diagrams and (d)–(f) PEV diagrams. On ROC diagrams, horizontal and vertical bars indicate the 95% confidence intervals as determined through blockwise bootstrapping. In PEV diagrams, 95% confidence intervals are indicated in transparent color backgrounds to mean PEV curves. On ROC diagrams, isolines of frequency bias B = 1 are drawn as blue solid and dashed lines for exact and relaxed timings, respectively; F is limited to the range of 0–0.35 for better visual separation of nearby points.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Comparison of performances of bias-corrected 12–36-h forecasts at wind farm EL for exact hourly timing of weather conditions and relaxed 6-hourly timing (“±3 h”), showing (a)–(c) ROC diagrams and (d)–(f) PEV diagrams. On ROC diagrams, horizontal and vertical bars indicate the 95% confidence intervals as determined through blockwise bootstrapping. In PEV diagrams, 95% confidence intervals are indicated in transparent color backgrounds to mean PEV curves. On ROC diagrams, isolines of frequency bias B = 1 are drawn as blue solid and dashed lines for exact and relaxed timings, respectively; F is limited to the range of 0–0.35 for better visual separation of nearby points.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Despite the positive effect of a relaxed timing constraint, the C/L range with positive PEV and the maximum PEV values remain limited, in particular in comparison to those for the other weather conditions. This begs the question if icing forecasts can be actually beneficial for users with real values of C/L. A detailed analysis of this aspect is beyond the scope of this study, since C/L depends on a number of wind farm site- and operation-specific factors. However, consider that the fixed selling price for electricity produced in EL is ~€103 (MW h)−1 (as of 2019) and the mean turbine power production is ~2.7 MW at a mean wintertime wind speed of ~9 m s−1. This gives a gross revenue of ~€278 h−1, coinciding with the loss L at turbine standstill. The blade heating of the turbine consumes a maximum of ~0.46 MW, leading to a reduction of revenue of ~€47 h−1 when heating is used, representing the cost C. This yields a C/L of ~0.17, which is well within the range of positive PEV (Fig. 10).
4) Comparison between sites
In the above, we have mostly focused on wind farm EL to discuss the need of bias corrections and tuning of ice growth rate thresholds and to demonstrate the effect of forecast timing. Figure 11 shows a comparison of forecast performances for EL and KH. Forecasts at both sites have been subjected to bias correction [section 5c(1)] using their respective two-winter error statistics. Overall, results at KH confirm those obtained at EL, when taking into account the differences in sample climatological rates. These differences are visible in the different positions of isolines for B on ROC diagrams and vertical lines at C/L = s on PEV diagrams. The background climatologies also result in clusters of points on the ROC diagram to be located in slightly different regions, since they give rise to different levels of false alarm rate F. For freezing and humid conditions (Figs. 11b,e), forecasts for KH tend to produce slightly lower PSS compared to those for EL, resulting in a lower and narrower PEV curve. For active ice growth (Figs. 11c,f), however, predictions at both locations are characterized by similar maximum PEV for the relaxed timing constraint, at a moderate degree of overforecasting. High-resolution forecasts perform better than low-resolution forecasts at both sites, if only by small margins. Generally, the good agreement between sites provides confidence in the robustness of our estimates of forecast skill.
Similar to Fig. 10, but for bias-corrected 12–36 h-forecasts at wind farms EL and KH. (c),(f) Verification results for active ice growth at the relaxed timing of ±3 h. On the ROC diagrams, isolines of frequency bias B = 1 and B = 1.5 are drawn as blue solid and dashed lines for EL and KH, respectively.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Similar to Fig. 10, but for bias-corrected 12–36 h-forecasts at wind farms EL and KH. (c),(f) Verification results for active ice growth at the relaxed timing of ±3 h. On the ROC diagrams, isolines of frequency bias B = 1 and B = 1.5 are drawn as blue solid and dashed lines for EL and KH, respectively.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Similar to Fig. 10, but for bias-corrected 12–36 h-forecasts at wind farms EL and KH. (c),(f) Verification results for active ice growth at the relaxed timing of ±3 h. On the ROC diagrams, isolines of frequency bias B = 1 and B = 1.5 are drawn as blue solid and dashed lines for EL and KH, respectively.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
5) Neighborhood ensemble forecasts
So far, single-point deterministic forecasts have been considered. Results have shown that icing forecasts have quite limited skill, in particular in comparison with forecasts of prerequisite conditions such as freezing temperature and high relative humidity. Here, we explore the potential of gridpoint neighborhood ensembles derived from WRF forecasts to represent the spatial uncertainty of icing forecasts. A similar approach was chosen by Molinder et al. (2018). Neighborhood ensemble forecasts and derived probabilities are constructed and verified as follows:
Forecasts are extracted from a volume of 3 × 3 × 3 = 27 grid points around the observation site. The resulting horizontal coverage is given by the model grid spacings of WRF D01 and WRF D02. The vertical extent of the neighborhood corresponds to the span of the rotor blades. Model data from the same heights above mean sea level are used for all grid points, which is considered a suitable procedure for the relatively smooth terrain around EL and KH.
Forecasts at each of the 27 points are corrected for their T, Td, and V biases, based on their two-winter forecast error statistics with respect to the observation point (section 5c).
Probabilities are derived from the 27-member neighborhood ensemble by counting the number of ensemble members predicting a given weather condition; all ensemble members are assumed to be equally likely (“frequentist” ensemble interpretation, e.g., Weigel 2012).
Pairs of hit rate and false alarm rate (H, F) are determined for probability forecasts, using a finite set of probability thresholds (>0%, 10%, 25%, 50%, 75%, and 90%).
PEV curves are computed individually for each probability threshold according to Eq. (6) and the optimal PEV curve is derived from the maximum PEV for each C/L (cf. Richardson 2012).
PEV diagrams of 12–36-h point forecasts and neighborhood ensemble forecasts at wind farms (a)–(c) EL and (d)–(f) KH for (left) weather condition T ≤ 0°C, (center) T ≤ 0°C and RH ≥ 85%, and (right) active ice growth at the relaxed timing of ±3 h.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
PEV diagrams of 12–36-h point forecasts and neighborhood ensemble forecasts at wind farms (a)–(c) EL and (d)–(f) KH for (left) weather condition T ≤ 0°C, (center) T ≤ 0°C and RH ≥ 85%, and (right) active ice growth at the relaxed timing of ±3 h.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
PEV diagrams of 12–36-h point forecasts and neighborhood ensemble forecasts at wind farms (a)–(c) EL and (d)–(f) KH for (left) weather condition T ≤ 0°C, (center) T ≤ 0°C and RH ≥ 85%, and (right) active ice growth at the relaxed timing of ±3 h.
Citation: Journal of Applied Meteorology and Climatology 59, 11; 10.1175/JAMC-D-20-0025.1
Interestingly, neighborhood forecasts derived from the coarse-resolution WRF D01 perform consistently better than the deterministic forecasts derived from the high-resolution WRF D02, the exception being the icing forecast for KH for C/L > s. Generally, however, the large confidence intervals surrounding mean PEV curves indicate limited statistical certainty. Longer verification periods and/or more sites would be needed to corroborate the results. Regardless, our findings underline the potential of probability icing forecasts, even when constructed with a neighborhood approach from a single forecast of coarse resolution. The improvement with respect to deterministic forecasts, however, is not great, pointing to the need of including other sources of uncertainty, such as in initial and boundary conditions and physics parameterizations.
6. Summary and conclusions
In this work, we have studied the skill of forecasts of atmospheric icing and related parameters for a two-winter period. The focus area has been a region in central Europe characterized by moderate terrain heights and icing climatologies of light and moderate severity. Observations and model forecasts at two wind farms (Ellern, Germany, and Kryštofovy Hamry, Czech Republic) are studied. Icing at these sites leads to significant losses in wind power production during winters, but the limited duration and severity of icing episodes would allow protective measures against ice-related turbine shutdown using anti-icing systems, provided warnings based on skillful forecasts are issued to wind farm operators ahead of time. The focus in our analysis is on the forecast range 12–36 h, with lead times from ~12 h being timely available to users in practice.
The skill of several different forecast models, ECMWF IFS, NCEP GFS, and WRF at coarse- and high-resolution is compared. The WRF Model setup was chosen to mimic that of operational forecasting systems (Benjamin et al. 2016) and of previous studies (e.g., Nygaard et al. 2011; Davis et al. 2014). Measurements of meteorological parameters and time series of active ice growth derived from camera image analyses are available for forecast verification at both wind farms.
Two case studies for the Ellern site serve to contrast two common synoptic patterns leading to icing: (i) a cold front associated with strong northwesterly flow, advecting a moist maritime air mass over central Europe, and (ii) a high pressure system with weak flow and widespread low stratus. The high-resolution WRF forecast accurately predicts ice accretion for the first case but fails for the second. A close inspection of the cases reveals possible sensitivities of the icing forecast to the modeled microphysical properties of the clouds/fog. However, explicit measurements of cloud water content and droplet size distributions would be needed to obtain further insight.
A two-winter verification study is conducted to assess the mean skill of forecasts and their potential economic value for users. Verification results are stratified in terms of forecast models, weather conditions, spatial and temporal resolution, and sites. The main findings are the following:
A regional-scale model verification of T, Td, and wind speed reveals a diurnal cycle in biases at the surface and, to a lesser degree, at turbine hub-height levels, especially for the WRF forecasts, possibly related to shortcomings in surface energy balance modeling. Biases as a function of forecast values are generally small for T but display marked positive trends for Td and wind speed, especially for WRF, making it more likely for WRF to produce icing if using the forecasts at face value.
A meaningful comparison of global and limited-area models for specific forecast sites is found to be possible only after bias correction of temperature, dewpoint temperature and wind speed. The skill of bias-corrected forecasts at predicting freezing temperature and high relative humidity is virtually identical for all models. This result confirms that the most critical aspect of icing forecasts is the modeling of liquid water content, with prerequisite freezing/saturated conditions being predicted comparatively much more easily.
To account for the different model horizontal resolutions and microphysical parameterizations, model-dependent optimal thresholds for ice growth rate are selected based on verification scores. The resulting icing forecasts for 12–36 h all tend to cluster on the ROC diagram with similar hit and false alarm rates and skill scores, with even the global models showing positive skill. Forecast performance at range 36–60 h generally is only marginally degraded compared to 12–36 h. Results strongly suggest the superiority of high-resolution WRF forecasts with respect to other model variants, although the margins of improvement are limited. In addition, broad 95% confidence intervals on verification scores point to the need for longer-term measurements and/or data from other wind farms to corroborate results.
No significant differences between WRF forecasts coupled to IFS and GFS are found, indicating that both global models can provide equally useful initial and boundary conditions for high-resolution icing forecasts.
The skill and potential economic value of icing forecasts at hourly temporal resolution is found to be quite limited (hit rates of ~0.40 and Peirce skill score of ~0.35). However, forecasts predicting the occurrence of icing within a period of, for example, six hours have substantially better accuracy (hit rates increased by ~0.2 or ~50%, Peirce skill score higher by ~0.15 or ~40%, leading to significantly higher potential economic value for a broader range of cost-loss ratios). Despite their coarser temporal resolution, such forecasts are deemed useful for users facing high losses from lost power production compared to relatively low-cost anti-icing procedures, such as blade heating. A brief back-of-the-envelope cost-loss analysis for wind farm Ellern lends credence to this conjecture [section 5c(3)]; however, a thorough case-by-case evaluation of user operational data, including revenue from sold energy and costs for heating, will be needed to determine the actual value for forecast users.
The potential for probability forecasts of icing has been explored through computationally inexpensive gridpoint neighborhood ensembles. Results clearly indicate added potential economic value compared to single-point forecasts, in particular at low cost-loss ratios. Even when derived from coarse-resolution neighborhood ensembles, probability forecasts outperform high-resolution deterministic forecasts, calling into question the need for high model resolution to produce icing forecasts.
The icing site climatologies considered here include only moderate icing severities. Analyses should be extended to other cold-climate or high-elevation sites, for instance, one of several existing wind farms in the Alps situated above 1500 m MSL.
The finding that the use of high-resolution forecasts offers only limited improvement with respect to low-resolution forecasts is in contrast to previous results, for example, to those by Nygaard et al. (2011) who evaluated explicitly the modeled LWC for cases of in-cloud icing and clearly demonstrated the benefits of grid spacings below 500 m. This is likely due to the much steeper terrain considered in those studies, in which case the dominant process of cloud formation was orographic lifting.
High-quality measurements of icing, suitable for forecast verification, are rare. In this study, icing data were available reliably only from camera image analyses; however, these did not allow to distinguish quantitatively between icing severities in the light to moderate range.
Last, this study was constrained to studying ice growth on a static structure (the turbine hub) rather than on-blade icing, which is ultimately responsible for degraded power production and turbine shutdown. Sensor systems explicitly measuring on-blade-accreted ice exist (cf. e.g., Meteotest 2016; Froidevaux et al. 2019) and some have just recently been exploited to establish a correlation of ice accretion between the nacelle and turbine blades (Jolin et al. 2019). For the present study, however, measurement data from such sensors were not available at sufficient quality and record length.
Acknowledgments
This research was funded by the Austrian Research Promotion Agency (FFG) within the second call of the “Energieforschungsprogramm des Klima und Energiefonds,” project ICE CONTROL (853575). A visit to NCAR by L. Strauss was supported through a stipend by the Ministry of Education, Science and Research of Austria. The computational results presented have been achieved using the Vienna Scientific Cluster (VSC). Access to ECMWF data was provided through the Zentralanstalt für Meteorologie und Geodynamik (ZAMG), Austria. Data for wind farm Kryštofovy Hamry were kindly made available by Enercon GmbH through a cooperation agreement. The authors thank Thomas Burchhart, Simon Kloiber, and all other members of the ICE CONTROL project for their individual contributions and are grateful to William Cheng, Gregory Thompson, Neil Davis, and Jennie Molinder for useful discussions. Last, the constructive criticism and suggestions for improvement by the editor and three anonymous reviewers are gratefully acknowledged.
Data availability statement
Access restrictions apply to some of the data used in this study. Surface and upper-air data and NCEP GFS forecasts are freely available, and sources have been cited in respective sections. ECMWF IFS forecasts are accessible to institutions of ECMWF member states. Access to measurement data from wind farm Ellern and WRF Model output produced during the ICE CONTROL project is limited to members of the project consortium, but selected data subsets can be made available by the authors upon request. Data from wind farm Kryštofovy Hamry, provided by Enercon GmbH, are not accessible to third parties.
REFERENCES
Arbez, C., M. Clément, C. Godreau, N. Swytink-Binnema, K. Tete, and M. Wadham-Gagnon, 2016: Development and validation of an ice prediction model for wind farms. Natural Resources Canada Tech. Rep., 134 pp., https://nergica.com/wp-content/uploads/3.NRCan_Report_EN.pdf.
Benjamin, S. G., and Coauthors, 2016: A North American hourly assimilation and model forecast cycle: The Rapid Refresh. Mon. Wea. Rev., 144, 1669–1694, https://doi.org/10.1175/MWR-D-15-0242.1.
Bergström, H., E. Olsson, S. Söderberg, P. Thorsson, and P. Undén, 2013: Wind power in cold climates: Ice mapping methods. Elforsk Rep. 13:10, 244 pp., http://www.diva-portal.org/smash/get/diva2:704372/FULLTEXT01.pdf.
Bernstein, B. C., J. Hirvonen, E. Gregow, and I. Wittmeyer, 2012: Experiences from real-time LAPS-LOWICE runs over Sweden: 2011-2012 icing season. Winterwind Int. Wind Energy Conf. 2012, Skellefteå, Sweden, Swedish Wind Power Association, http://www.slideshare.net/WinterwindConference/3a-bernstein-lapslowice.
Bredesen, R., M. Drapalik, and B. Butt, 2017a: Understanding and acknowledging the ice throw hazard—Consequences for regulatory frameworks, risk perception and risk communication. J. Phys.: Conf. Ser., 926, 012001, https://doi.org/10.1088/1742-6596/926/1/012001.
Bredesen, R., and Coauthors, 2017b: Wind energy projects in cold climates. IEA Wind TCP Task 19 Rep., 2nd ed. 49 pp., https://community.ieawind.org/task19/viewdocument/iea-wind-tcp-task-19-recommended-pr.
Colarco, P., A. da Silva, M. Chin, and T. Diehl, 2010: Online simulations of global aerosol distributions in the NASA GEOS-4 model and comparisons to satellite and ground-based aerosol optical depth. J. Geophys. Res., 115, D14207, https://doi.org/10.1029/2009jd012820.
Combitech, 2016: IceMonitor product sheet. Combitech, 2 pp., accessed 27 July 2020, http://www.rwis.net/res/pdffiles/IceMonitor_Product_Sheet.pdf.
Davis, N., A. N. Hahmann, N.-E. Clausen, and M. Žagar, 2014: Forecast of icing events at a wind farm in Sweden. J. Appl. Meteor. Climatol., 53, 262–281, https://doi.org/10.1175/JAMC-D-13-09.1.
Davis, N., Ø. Byrkjedal, A. N. Hahmann, N.-E. Clausen, and M. Žagar, 2016a: Ice detection on wind turbines using the observed power curve. Wind Energy, 19, 999–1010, https://doi.org/10.1002/we.1878.
Davis, N., P. Pinson, A. N. Hahmann, N.-E. Clausen, and M. Žagar, 2016b: Identifying and characterizing the impact of turbine icing on wind farm power generation. Wind Energy, 19, 1503–1518, https://doi.org/10.1002/we.1933.
Déqué, M., 2012: Deterministic forecasts of continuous variables. Forecast Verification: A Practitioner’s Guide in Atmospheric Science, 2nd ed. I. T. Jolliffe and D. B. Stephenson, Eds., John Wiley & Sons, 77–94, https://doi.org/10.1002/9781119960003.
Drage, M. A., and G. Hauge, 2008: Atmospheric icing in a coastal mountainous terrain. Measurements and numerical simulations, a case study. Cold Reg. Sci. Technol., 53, 150–161, https://doi.org/10.1016/j.coldregions.2007.12.003.
ECMWF, 2020: IFS documentation. Accessed 27 July 2020, https://www.ecmwf.int/en/publications/ifs-documentation.
Forbes, R., A. Tompkins, and A. Untch, 2011: A new prognostic bulk microphysics scheme for the IFS. ECMWF Tech. Memo. 649, 22 pp., https://www.ecmwf.int/en/elibrary/9441-new-prognostic-bulk-microphysics-scheme-ifs.
Froidevaux, P., S. Bourgeois, and R. Cattin, 2019: Intercomparison of blade-based ice detection systems. Winterwind Int. Wind Energy Conf. 2019, Umeå, Sweden, Winterwind, https://winterwind.se/wp-content/uploads/2019/02/09_03_Froidevaux_Benchmark_of_four_Blade-based_Ice_Detection_Systems_Pub_v2.pdf.
Ginoux, P., M. Chin, I. Tegen, J. Prospero, B. Holben, O. Dubovik, and S. Lin, 2001: Sources and distributions of dust aerosols simulated with the GOCART model. J. Geophys. Res., 106, 20 255–20 273, https://doi.org/10.1029/2000JD000053.
Grell, G. A., and S. R. Freitas, 2014: A scale and aerosol aware stochastic convective parameterization for weather and air quality modeling. Atmos. Chem. Phys., 14, 5233–5250, https://doi.org/10.5194/acp-14-5233-2014.
Hogan, R. J., and I. B. Mason, 2012: Deterministic forecasts of binary events. Forecast Verification: A Practitioner’s Guide in Atmospheric Science, 2nd ed. I. T. Jolliffe and D. B. Stephenson, Eds., John Wiley and Sons, 31–59, https://doi.org/10.1002/9781119960003.
Iacono, M. J., J. S. Delamere, E. J. Mlawer, M. W. Shephard, S. A. Clough, and W. D. Collins, 2008: Radiative forcing by long-lived greenhouse gases: Calculations with the AER radiative transfer models. J. Geophys. Res., 113, D13103, https://doi.org/10.1029/2008JD009944.
ISO, 2017: Atmospheric icing of structures. ISO 12494:2017(en), https://www.iso.org/obp/ui#iso:std:iso:12494:ed-2:v1:en.
Jokela, T., M. Tiihonen, and T. Karlsson, 2019: Validation of droplet size in the VTT icing wind tunnel test section. Winterwind Int. Wind Energy Conf. 2019, Umeå, Sweden, Winterwind, 12 pp., https://winterwind.se/wp-content/uploads/2019/02/11_01_Jokela_Validation_of_Droplet_Size_in_the_VTT_Icing_Wind_Tunnel_Test_Section_Pub_v1.pdf.
Jolin, N., D. Bolduc, N. Swytink-Binnema, G. Rosso, and C. Godreau, 2019: Wind turbine blade ice accretion: A correlation with nacelle ice accretion. Cold Reg. Sci. Technol., 157, 235–241, https://doi.org/10.1016/j.coldregions.2018.10.009.
Kaikkonen, V. A., E. O. Molkoselka, and A. J. Makynen, 2020: A rotating holographic imager for stationary cloud droplet and ice crystal measurements. Opt. Rev., 27, 205–216, https://doi.org/10.1007/s10043-020-00583-y.
Klemp, J. B., J. Dudhia, and A. D. Hassiotis, 2008: An upper gravity-wave absorbing layer for NWP applications. Mon. Wea. Rev., 136, 3987–4004, https://doi.org/10.1175/2008MWR2596.1.
Krenn, A., and Coauthors, 2018: International recommendations for ice fall and ice throw risk assessments. IEA Wind TCP Task 19 Rep., 41 pp., https://community.ieawind.org/HigherLogic/System/DownloadDocumentFile.ashx?DocumentFileKey=3e92fc30-a54a-4888-e612-79126301c58e&forceDialog=1.
Lehtomäki, V., 2016: Emerging from the cold. Wind Power Mon., 29 July, https://www.windpowermonthly.com/article/1403504/emerging-cold.
Lehtomäki, V., and Coauthors, 2018: Available technologies for wind energy in cold climates. IEA Wind TCP Task 19 Rep., 129 pp., https://community.ieawind.org/task19/viewdocument/available-technologies-for-wind-ene.
Makkonen, L., 2000: Models for the growth of rime, glaze, icicles and wet snow on structures. Philos. Trans. Roy. Soc., A358, 2913–2939, https://doi.org/10.1098/rsta.2000.0690.
Meteotest, 2016: Evaluation of ice detection systems for wind turbines. VGB Research Project 392 Final Rep., 111 pp., https://www.vgb.org/vgbmultimedia/392_Final+report-p-10476.pdf.
Molinder, J., H. Körnich, E. Olsson, H. Bergström, and A. Sjöblom, 2018: Probabilistic forecasting of wind power production losses in cold climates: A case study. Wind Energy Sci., 3, 667–680, https://doi.org/10.5194/wes-3-667-2018.
Molinder, J., H. Körnich, E. Olsson, and P. Hessling, 2019: The use of uncertainty quantification for the empirical modeling of wind turbine icing. J. Appl. Meteor. Climatol., 58, 2019–2032, https://doi.org/10.1175/JAMC-D-18-0160.1.
Nakanishi, M., and H. Niino, 2004: An improved Mellor-Yamada level-3 model with condensation physics: Its design and verification. Bound.-Layer Meteor., 112, 1–31, https://doi.org/10.1023/B:BOUN.0000020164.04146.98.
Nakanishi, M., and H. Niino, 2006: An improved Mellor-Yamada level-3 model: Its numerical stability and application to a regional prediction of advection fog. Bound.-Layer Meteor., 119, 397–407, https://doi.org/10.1007/s10546-005-9030-8.
NCEP/NOAA, 2020: Global Forecast System. Accessed 27 July 2020, https://www.emc.ncep.noaa.gov/emc/pages/numerical_forecast_systems/gfs.php.
NCEP/NWS/NOAA/U.S. Department of Commerce, 2008: NCEP ADP global upper air and surface weather observations (PREPBUFR format). National Center for Atmospheric Research Computational and Information Systems Laboratory Research Data Archive, accessed 27 July 2020, https://doi.org/10.5065/Z83F-N512.
Nygaard, B. E., 2009: Evaluation of icing simulations for “COST727 icing test sites” in Europe. 13th Int. Workshop on Atmospheric Icing of Structures (IWAIS XIII), Andermatt, Switzerland, IWAIS, https://www.compusult.com/html/IWAIS_Proceedings/IWAIS_2009/Session_3_cost_727_wg1/session_3_nygaard.pdf.
Nygaard, B. E. K., J. E. Kristjánsson, and L. Makkonen, 2011: Prediction of in-cloud icing conditions at ground level using the WRF Model. J. Appl. Meteor. Climatol., 50, 2445–2459, https://doi.org/10.1175/JAMC-D-11-054.1.
Nygaard, B. E. K., H. Ágústsson, and K. Somfalvi-Toth, 2013: Modeling wet snow accretion on power lines: Improvements to previous methods using 50 years of observations. J. Appl. Meteor. Climatol., 52, 2189–2203, https://doi.org/10.1175/JAMC-D-12-0332.1.
Podolskiy, E. A., B. E. K. Nygaard, K. Nishimura, L. Makkonen, and E. P. Lozowski, 2012: Study of unusual atmospheric icing at Mount Zao, Japan, using the weather research and forecasting model. J. Geophys. Res., 117, D12106, https://doi.org/10.1029/2011JD017042.
Richardson, D. S., 2012: Economic value and skill. Forecast Verification: A Practitioner’s Guide in Atmospheric Science, 2nd ed. I. T. Jolliffe and D. B. Stephenson, Eds., John Wiley and Sons, 167–184, https://doi.org/10.1002/9781119960003.
Rydblom, S., B. Thornberg, and E. Olsson, 2019: Field study of LWC and MVD using the droplet imaging instrument. IEEE Trans. Instrum. Meas., 68, 614–622, https://doi.org/10.1109/TIM.2018.2843599.
Scher, S., and J. Molinder, 2019: Machine learning-based prediction of icing-related wind power production loss. IEEE Access, 7, 129 421–129 429, https://doi.org/10.1109/ACCESS.2019.2939657.
Skamarock, W. C., and J. B. Klemp, 2008: A time-split nonhydrostatic atmospheric model for weather research and forecasting applications. J. Comput. Phys., 227, 3465–3485, https://doi.org/10.1016/j.jcp.2007.01.037.
Smirnova, T. G., J. M. Brown, S. G. Benjamin, and J. S. Kenyon, 2016: Modifications to the Rapid Update Cycle Land Surface Model (RUC LSM) available in the Weather Research and Forecasting (WRF) Model. Mon. Wea. Rev., 144, 1851–1865, https://doi.org/10.1175/MWR-D-15-0198.1.
Smith, A., N. Lott, and R. Vose, 2011: The integrated surface database: Recent developments and partnerships. Bull. Amer. Meteor. Soc., 92, 704–708, https://doi.org/10.1175/2011BAMS3015.1.
Sundqvist, H., E. Berge, and J. E. Kristjansson, 1989: Condensation and cloud parameterization studies with a mesoscale numerical weather prediction model. Mon. Wea. Rev., 117, 1641–1657, https://doi.org/10.1175/1520-0493(1989)117<1641:CACPSW>2.0.CO;2.
Thompson, G., and T. Eidhammer, 2014: A study of aerosol impacts on clouds and precipitation development in a large winter cyclone. J. Atmos. Sci., 71, 3636–3658, https://doi.org/10.1175/JAS-D-13-0305.1.
Thompson, G., P. R. Field, R. M. Rasmussen, and W. D. Hall, 2008: Explicit forecasts of winter precipitation using an improved bulk microphysics scheme. Part II: Implementation of a new snow parameterization. Mon. Wea. Rev., 136, 5095–5115, https://doi.org/10.1175/2008MWR2387.1.
Thompson, G., B. E. Nygaard, L. Makkonen, and S. Dierer, 2009: Using the Weather Research and Forecasting (WRF) model to predict ground/structural icing. 13th Int. Workshop on Atmospheric Icing of Structures (IWAIS XIII), Andermatt, Switzerland, IWAIS, https://www.compusult.com/html/IWAIS_Proceedings/IWAIS_2009/Session_3_cost_727_wg1/session_3_thompson.pdf.
Thorsson, P., S. Söderberg, and H. Bergström, 2015: Modelling atmospheric icing: A comparison between icing calculated with measured meteorological data and NWP data. Cold Reg. Sci. Technol., 119, 124–131, https://doi.org/10.1016/j.coldregions.2015.07.003.
Vassbø, T., J. Kristjansson, S. Fikke, and L. Makkonen, 1998: An investigation of the feasibility of predicting icing episodes using numerical weather prediction model output. Proc. Eighth Int. Workshop on Atmospheric Icing of Structures (IWAIS VIII), Reykjavik, Iceland, IWAIS, 343–347.
Weigel, A. P., 2012: Ensemble forecasts. Forecast Verification: A Practitioner’s Guide in Atmospheric Science, 2nd ed. I. T. Jolliffe and D. B. Stephenson, Eds., John Wiley and Sons, 141–166, https://doi.org/10.1002/9781119960003.
Weissinger, M., 2017: Synoptische Analyse von Vereisungsfällen an Windkraftanlagen am Beispiel Ellern, Deutschland (Synoptic analysis of icing events at wind turbines in Ellern, Germany). Bachelor’s thesis, University of Vienna, 53 pp.
Wilks, D., 1997: Resampling hypothesis tests for autocorrelated fields. J. Climate, 10, 65–82, https://doi.org/10.1175/1520-0442(1997)010<0065:RHTFAF>2.0.CO;2.
Wilks, D., 2011: Statistical Methods in the Atmospheric Sciences. 3rd ed. Elsevier, 676 pp.
Yang, J., K. F. Jones, W. Yu, and R. Morris, 2012: Simulation of in-cloud icing events on Mount Washington with the GEM-LAM. J. Geophys. Res., 117, D17204, https://doi.org/10.1029/2012JD017520.
Zhao, Q., and F. H. Carr, 1997: A prognostic cloud scheme for operational NWP models. Mon. Wea. Rev., 125, 1931–1953, https://doi.org/10.1175/1520-0493(1997)125<1931:APCSFO>2.0.CO;2.