Abstract

The wind power industry is highly sensitive to weather, and there is a clear impact on turbine efficiency associated with icing, which can cause significant power losses and even result in the total shutdown of wind farms. Therefore, accurate analyses and forecasts of wind- and icing-related meteorological variables are of great importance. To this end, the Local Analysis and Prediction System (LAPS)–LOWICE system has been developed to produce real-time, hourly estimates of the presence, intensity, and impacts of icing on wind power production. As part of this development, it became clear that power losses did not correlate well with measured icing loads but correlated reasonably well with the time history of icing rate in combination with ice loss due to melting, sublimation, and shedding.

1. Introduction

The expanding use of renewable energy implies new developments within the wind power industry, and the global market for wind energy grew more than 56 billion Euro during 2011–12 (GWEC 2012). In Europe, wind power is now the fastest-growing source of energy (EWEA 2013). The energy available from wind is highly variable in time and space; therefore, better wind prediction may allow power companies to determine how, when, and where to transfer the energy generated into the electricity network. It is important to recognize the large economic impact such decisions can make. For example, the buying and selling of electricity is highly dependent on the forecasts of the potential for wind energy that is expected in the coming hours and days. Beyond the direct and obvious effects of wind speed, the accretion and persistence of ice can have a large impact on turbine efficiency and thus the amount of electricity generated. In some cases, the effects of icing from supercooled clouds, freezing precipitation, and wet snow can result in complete power loss and even shutdown of wind farms (Tammelin et al. 2000). Icing can also cause dynamic instability on a turbine due to irregular distribution of ice, resulting in fatigue and faster wear of mechanical parts. Furthermore, accumulated ice can fall from, or be thrown from, the turbine when it detaches from the blades and other surfaces, posing a potential risk of injuries to humans and nearby structures (Cattin et al. 2007). These elements of wind power are becoming increasingly important as more wind power capacity is being built in cold climates, where there is a large threat of icing.

Accurate analysis and forecasting of wind speed and icing can allow meteorologists to provide power companies and traders with valuable estimates of expected power and can allow those controlling the turbines to make informed decisions about when to activate deicing equipment and/or to shut down turbines to reduce the risks of damage to the turbines and their surroundings. Thus, there is a clear need for timely, high-quality analyses and forecasts of wind power and the effects of icing thereon.

It is with these concepts in mind that a wind power icing diagnostic system has been developed for Scandinavia, combining output from the Finnish Meteorological Institute’s Local Analysis and Prediction System (LAPS: Albers et al. 1996) model and the LOWICE algorithm (Bernstein et al. 2011a). LAPS–LOWICE produces real-time, hourly estimates of wind speed and the presence, rate, and impacts of icing on wind power production. LOWICE employs multiple approaches to estimate the proportion of power expected to be lost due to icing. The first, simplest method [version 0 (V0)] ties power loss solely to the amount of ice “load,” while newer versions [version 1 (V1) and version 2 (V2)] tie power loss to the recent history of icing rate, as well as the return of power due to melting, sublimation, and shedding of previously accreted ice. In this paper, LAPS–LOWICE is described and the output will be validated and verified against observations from a wind farm in Sweden.

2. Methods and material

Estimation of the likelihood and intensity of near-surface icing conditions is a challenging problem for meteorologists and those affected by this phenomenon, including the power industry. Beyond direct measurements of icing, wind, and power at a given location, there are numerous sources of independent meteorological data that can be used to estimate near-surface icing conditions indirectly. In particular, observations from satellites, radiosondes, surface stations, and radars (if available) provide a great deal of useful information, especially when paired with forecasts from numerical weather models. Each of these data sources has its strengths and weaknesses for the analysis and forecasting of icing conditions, and the information from each must be considered carefully in the context of the meteorological environment if information from them is to be employed effectively (as in Tafferner et al. 2003; Le Bot 2004). By blending 3D numerical model fields and observations from sources such as those described above in a manner that is consistent with the meteorology of icing, LAPS–LOWICE is able to produce high-resolution grids of icing probability, icing rate, and ice load, as well as expected “clean” (ice free) and “iced” power across Scandinavia.

a. LAPS

The Finnish Meteorological Institute (FMI) operates the LAPS to produce 3D analyses of a number of meteorological parameters (Albers et al. 1996; Koskinen et al. 2011). LAPS uses a data fusion method to generate high-resolution analyses using statistical methods to merge coarser-resolution background fields. Observations are fitted to the coarser first-guess analysis mainly by successive correction, while high-resolution topographical datasets are taken into account when creating the final analysis fields. Those analysis products are mainly used for nowcasting purposes, that is, what is currently happening and what will happen in the next few hours.

In FMI-LAPS, background first-guess fields are derived from the European Centre for Medium-Range Weather Forecasts (ECMWF) model grids, with horizontal grid spacing of ~16 km (ECMWF 2014). This includes 3D forecasts of geopotential height, temperature, specific humidity, and winds at 16 pressure levels. Single-level variables include surface geopotential height, surface pressure, 2-m temperature, and 10-m wind. FMI-LAPS output is produced at 3-km horizontal spacing across Scandinavia (Fig. 1) and on 44 vertical (pressure) levels, with 10-hPa spacing near the surface. LAPS relies heavily on high-resolution data from remote sensors and ground-based observational networks, including Meteosat satellite, surface synoptic observations (SYNOPs) and METARs, precipitation gauges, road weather sensors, soundings, AMDAR temperatures and winds, and radar data. LAPS output fields include 3D pressure, geopotential height, temperature, winds, relative humidity, model condensate, and both 2D and 3D cloud fields [e.g., cloud-top temperature, cloud-top and cloud-base height, fractional cloud cover, and cloud layering; described in more details in section 2a(1)]. Many of these LAPS fields are highly relevant for both wind power and icing, providing essential inputs to the LOWICE system.

Fig. 1.

(a) Areal coverage of LAPS analysis domain is shown with a black line. (b) Surface observations (black crosses) and sounding stations (open black circles) ingested into LAPS, together with the topography height (see color bar). The wind farm is located in the mountains of western Sweden, at an elevation of approximately 800 m above mean sea level. The location is within the enclosed area (red dotted line) in (b), which describes the region SE2 of the Swedish electrical power net.

Fig. 1.

(a) Areal coverage of LAPS analysis domain is shown with a black line. (b) Surface observations (black crosses) and sounding stations (open black circles) ingested into LAPS, together with the topography height (see color bar). The wind farm is located in the mountains of western Sweden, at an elevation of approximately 800 m above mean sea level. The location is within the enclosed area (red dotted line) in (b), which describes the region SE2 of the Swedish electrical power net.

LAPS cloud analysis

One of the dominant causes of icing on wind power plants is the freezing of supercooled liquid cloud drops onto turbine blades. To determine the presence or absence of clouds (the “cloud mask”), their heights, layering, etc., LAPS relies heavily on satellite data from Meteosat. During daytime, visible and multiple infrared (IR) channels can be used to determine the cloud mask, while only infrared channels are used at night. To estimate the cloud-top height, longwave IR temperatures are compared to the LAPS temperature and height profiles. LAPS then assimilates other existing measurements, such as surface observations (METARs and SYNOPs) to assess the cloud-base height (Fig. 2).

Fig. 2.

FMI-LAPS cloud analysis flow diagram, describing the sequential processes of assimilating different input data, after Albers et al. (1996).

Fig. 2.

FMI-LAPS cloud analysis flow diagram, describing the sequential processes of assimilating different input data, after Albers et al. (1996).

In northern latitudes, there is a high frequency of low clouds due to frontal systems, maritime influences, and temperature inversions. Especially during winter, these single-layered low clouds form due to strong temperature inversion, where the condensate is trapped and potentially induces a risk of icing. In these situations, it has been common for cloud-top heights to be overestimated (as in Haggerty et al. 2008). LAPS searches the temperature profile downward, in order to match it with the measured cloud-top IR temperature, and the cloud top is put at the first vertical level where these temperatures are found. When inversions are present, multiple levels may meet this criterion. However, this was not considered in the original LAPS cloud-top identification method. Therefore, new algorithms have been added to locate subinversion clouds in weather situations where only a single-layer low cloud is present. When the new method finds a temperature inversion, the cloud height is set below the thermal inversion if the satellite infrared temperature is greater than the air temperature below the inversion by at least 10°C. As a result, the vertical cloud structure can be significantly changed in weather situations with strong low-level inversions. This cloud inversion method follows the same principles as those applied to Meteosat data by EUMETSAT (SAFNWC 2013).

In addition, the FMI-LAPS cloud analysis (Fig. 2) now uses cloud mask products from the Nowcasting Satellite Application Facilities at EUMETSAT (SAFNWC) to correct the cloud mask in the final stage of cloud analysis (Derrien and Le Gléau 2005). There are certain problems within the cloud analysis that are very difficult to solve, such as twilight effects (dawn/dusk), resulting in weak information from satellite channels, and how to distinguish between open and frozen water bodies, which sometimes cause overestimations of clouds over sea/lakes, especially during nighttime (Dybbroe et al. 2005).

b. LOWICE

The basic concepts behind LOWICE are similar to those used in the current icing product (CIP; Bernstein et al. 2005, 2006), which combines numerical model output with observations from satellite, surface stations, pilot reports, and a lightning detection network to produce 3D analyses of the probability and intensity of icing conditions aloft over the contiguous United States for the aviation community. LOWICE includes numerous unique concepts and code that are specific to the near-surface icing environment. In sections 2b(1)2b(3) the LOWICE processes used to estimate wind power and power loss due to icing will be described.

1) Icing detection, likelihood, and intensity

Using the meteorological fields described above, LOWICE examines each vertical column of LAPS data. First, if clouds are present, then the vertical profiles of cloud and relative humidity (RH) are examined, along with observations of precipitation from surface stations, to determine whether the following environments appear to be present: single-layer clouds, multilayer clouds, classical freezing rain, nonclassical freezing drizzle (Huffman and Norman 1988), or a snow-dominated precipitation process. For each of these environments, the column of data must be examined uniquely to correctly assess the expected likelihood and intensity of icing near the surface, where wind turbines operate. Icing is considered possible at the turbine level if the temperature is subfreezing and the level is located between the highest cloud top and the lowest cloud base. Levels just below cloud base are also allowed to have icing due to uncertainty in ceiling height, especially in complex terrain. In addition, the possibility for icing is considered for all subfreezing levels between cloud base and the surface when liquid or freezing precipitation is reported, since these observations imply the potential for freezing precipitation above the ground when certain meteorological structures are present (e.g., the classical freezing rain structure). A flowchart showing the LOWICE analysis process is given in Fig. 3.

Fig. 3.

LOWICE icing and power assessment flow diagram.

Fig. 3.

LOWICE icing and power assessment flow diagram.

Cloud-top temperature (CTT) is a very useful parameter for assessing the likelihood of icing at a given location. Therefore, it is important to determine whether the clouds exposed to satellite view are those that are directly affecting the wind farm (Fig. 4). LOWICE performs its own assessment of the presence of multiple cloud layers, following the simple examination of the RH vertical profile described in Bernstein et al. (2005), searching downward from the highest cloud tops, looking for layers with RH ≤ 50%. If at least three model levels meet this criterion, then a dry layer is considered to be present. The downward progression continues and the system searches for a layer with RH ≥ 70%. If that criterion is met, then a new cloud layer is identified and the CTT of that layer is set to the temperature (T) where RH ≥ 70% was found. This process continues until the lowest cloud base height (CBZ) is reached, and then the CTT of the layer affecting the site (CTT-farm) is set to the CTT of the lowest cloud layer found above the site.

Fig. 4.

Conceptual model of the LOWICE multilayer cloud scheme. (left) The RH profile is shown. (right) The CTT-sat and CTT-farm represents the CTT of the upper and lower cloud layers (gray), respectively. The horizontal solid line represents the elevation of the site (i.e., turbine hub height).

Fig. 4.

Conceptual model of the LOWICE multilayer cloud scheme. (left) The RH profile is shown. (right) The CTT-sat and CTT-farm represents the CTT of the upper and lower cloud layers (gray), respectively. The horizontal solid line represents the elevation of the site (i.e., turbine hub height).

The icing likelihood [ICElike; Eq. (1)] is then initially estimated by multiplying the output from LOWICE’s membership functions for T, RH, and CTT at the hub height (Tmap, RHmap, CTTmap; Fig. 5; Bernstein et al. 2005). These functions are designed to account for some of the uncertainty inherent in the data being used, as well as the physical characteristics of supercooled liquid water in the atmosphere (e.g., icing is more likely to exist at relatively warm, subfreezing temperatures). ICElike is then increased or decreased based upon reports of certain precipitation types (e.g., freezing drizzle, freezing fog, snow, snow grains) that have implications about icing. Nearby observations have a much greater influence than those that are more distant,

 
formula

To estimate the icing rate, it is necessary to estimate T, liquid water content (LWC), and wind speed (υ) at the site. Here T and υ are taken directly from LAPS data, vertically interpolated to the site elevation. LOWICE then estimates LWC using the cloud and precipitation fields described above in combination with LAPS profiles of pressure (P), T, and RH. If the level of interest is located at or above the estimated CBZ, then a first-guess LWC is generated by assuming a moist adiabatic lapse rate from the level of interest down to CBZ. This is done by estimating the saturated mixing ratio at CBZ, via standard equations, and taking the difference between the mixing ratios at CBZ and the height of the wind farm (Z-farm), and then compensating for density. While using the adiabatic assumption is not ideal, it is not unusual for lapse rates to be nearly moist adiabatic over the shallow layer between Z-farm and CBZ. When lapse rates are more stable than moist adiabatic, initial LWC estimates are decreased to be more realistic. As noted earlier, because of the potential for local variability in CBZ between the site and the nearest station-reported ceiling height, LOWICE also allows for the possibility for icing at elevations slightly below CBZ. This allowable depth beneath the cloud base (dz) changes with the distance between the site and the reporting surface station. It is at its minimum value when the ceiling report is made close to the site and increases linearly when the ceiling report is made at the maximum allowable distance of 160 km. LWC estimates are set to nominal values between 0.0 and 0.1 g m−3 as dz increases from zero to the maximum allowable value.

Fig. 5.

Membership functions for (a) temperature (Tmap, gray line) and CTT (CTTmap, black line), and (b) RH (RHmap, %, black line). The x axes are temperature (°C) and RH with respect to water (%), while the y axes are unitless.

Fig. 5.

Membership functions for (a) temperature (Tmap, gray line) and CTT (CTTmap, black line), and (b) RH (RHmap, %, black line). The x axes are temperature (°C) and RH with respect to water (%), while the y axes are unitless.

One important aspect of the adiabatic assumption is that all condensate remains within the cloud. Many icing clouds produce precipitation, which depletes the supercooled LWC (SLWC) within them. This is particularly true when snow is expected to be present within the clouds, because the SLWC is expected to be at least partially depleted via riming (Rogers and Yau 1989). The closer the observation of precipitation is to the site, the greater the likelihood that depletion is occurring at the site. To address this process, LOWICE examines all surface observations within the radius of influence for the occurrence of precipitation. If precipitation is observed, then a depletion factor is calculated based on the distance between the site and the precipitation report. The initial estimate of LWC can be depleted by as much as 50%, depending on the distance to the reporting station. This approach is reasonably consistent with evidence from flights made in precipitating, well-mixed icing clouds sampled during natural icing flight programs (e.g., Bernstein et al. 2011b; Politovich and Bernstein 1995). Radar data can also be quite valuable for estimating the depletion of SLWC by riming, but radar data were not readily available over Sweden in LAPS through 2012 and therefore were not used in this version of LOWICE.

Once the T, LWC, and υ have been estimated, the next step is to calculate the icing rate (ICErate). This is done using a standard icing-rate equation for a cylinder (Finstad et al. 1988; Makkonen 2000; ISO 2001):

 
formula

where SLWC is the supercooled LWC (g m−3); A is the cross-sectional area (m2) of the object; υ is the wind velocity (m s−1); and the unitless η1, η2, and η3 are the correction factors for the collision, sticking, and accretion efficiencies, respectively, which are all set to 1.0 for simplicity. In this case, A is set equal to 0.015 m2, based on the International Organization for Standardization (ISO) 12494 standard reference cylinder, which is 0.5 m long and has a 30-mm diameter. Because dm/dt is in grams per second and LOWICE values are produced once per hour, dm/dt is multiplied by 3600 (s h−1) to generate the hourly icing rate in grams per hour. The final equation for the icing rate takes on the form

 
formula

2) Ice-load and ice-loss mechanisms

Applying the icing rate (ICErate) for each hour, ice is accumulated and the ice mass (ICEmass) on the reference cylinder increases accordingly. ICEmass builds when icing is active (ICErate > 0 kg h−1) and is depleted by melting and sublimation when icing is not active (ICErate = 0 kg h−1):

 
formula
 
formula

where the subscripts “0” and “1” indicate the values from the previous and current times, respectively. Note that ice load (N) is calculated by multiplying ice mass (kg) by gravitational acceleration (~9.81 m s−2). Melting of ice loads is obviously important, since even large ice loads can dissipate fairly quickly once the temperature becomes adequately warm. Rather than simply melting away the entire ice load as soon as the model temperature (T) exceeds 0°C, it is logical that a melting rate should be applied, based on how much temperature exceeds 0°C. Thus, LOWICE estimates a simple melting rate (MELTrate), which increases linearly from 0.0 kg h−1 (when T ≤ 0°C) to a maximum value of 10.0 kg h−1 (when T ≥ +5°C). This estimation of melting rate is somewhat arbitrary, but sensitivity tests performed on several years of historical icing events at wind farms indicated that changes in the melting rate parameter had little effect on the overall estimates of ice load. Future versions of LOWICE may also include insolation as a factor, since sunlight on turbine blades could increase their temperature above 0°C, even when the air temperature is below 0°C.

Sublimation also plays an important role in the depletion of ice loads, especially in the Scandinavian winter, when temperatures rarely exceed 0°C. Evidence from the examination of meteorological measurements and webcam images during subfreezing, dry, windy weather indicated that the visible ice (on webcam images) and the measured loads decreased gradually over time (e.g., Fig. 6). In this figure, notice that visibility (gray line) was generally low during the period of active icing and then high during the period of sublimation. Such changes in visibility are often associated with the presence and absence of clouds that are associated with icing (as in Portin et al. 2009). LOWICE’s simple sublimation scheme was developed empirically, based on a combination of wind speed and relative humidity [see Eq. (4c)]. Naturally, sublimation occurs more quickly when winds speed (υ) is large in subsaturated environments and, to a lesser extent, when RH is low:

 
formula

where SUBrate is in kilograms per hour, υ is in meters per second, and RHmap is a fuzzy logic membership function based on RH (see Fig. 5). Equation (4c) is weighted more heavily toward v than RH. Because RHmap = 0.0 for RH ≤ 70% and RHmap = 1.0 for RH ≥ 90%, thereby the RH part of the equation has no effect when RH ≥ 90% and a relatively strong effect when RH ≤ 70%. The entire equation is multiplied by an ice-load change rate of 0.2 kg h−1, so the most ice that the system can sublimate in an hour is 0.2 kg (when υ ≥ 10 m s−1 and RH ≤ 70%). This rate may seem insubstantial, but the effects can be dramatic, especially when periods of icing are followed by extended periods of dry, windy, subfreezing weather. During such periods, ice cannot be removed by melting but sublimation can gradually deplete the ice (and its effects) over time. When icing is inactive (ICErate = 0 kg h−1), ICEmass is decreased using the sublimation and melting equations described above. Without question, the sublimation and melting schemes presented here are an oversimplification of complex processes. The use of more complex melting and sublimation schemes from published literature should be tested in the future.

Fig. 6.

Time series plot to illustrate the effects of sublimation. Observations from one wind farm in Sweden during periods of active icing (before 1100 UTC 7 Jan 2011) and sublimation (after 1100 UTC 7 Jan 2011). Observations are plotted against the same y axis, with different units; temperature (air temperature, °C), wind speed at 10 m (wind speed, m s−1), ice load [ice load, N (0.5 m)−1], RH (RH/10, %, divided by 10 for plotting purposes), and visibility (visibility, m).

Fig. 6.

Time series plot to illustrate the effects of sublimation. Observations from one wind farm in Sweden during periods of active icing (before 1100 UTC 7 Jan 2011) and sublimation (after 1100 UTC 7 Jan 2011). Observations are plotted against the same y axis, with different units; temperature (air temperature, °C), wind speed at 10 m (wind speed, m s−1), ice load [ice load, N (0.5 m)−1], RH (RH/10, %, divided by 10 for plotting purposes), and visibility (visibility, m).

Another process that is important for removing ice from a wind turbine is the shedding process. Shedding tends to be stochastic in nature, with sudden events and effects that are very difficult to predict. Anecdotal evidence suggests that shedding occurs primarily during periods of melting but also during periods of high wind speeds, regardless of temperature. When blades are turning, increases in wind speed dramatically increase the centrifugal force on the blades, especially toward the blade tips, and can rapidly become strong enough to cause the ice to de-bond and shed.

3) Wind power estimations at turbine—LOWICE V0, V1, and V2

To estimate the potential impacts on power production, the power that the turbines are expected to generate when they are free of ice must first be estimated. To this end, measured wind speeds and power output were compared over ice-free portions of autumn at several wind farms. Using 0.1 m s−1 wind speed bins, a scatterplot was created, then a piecewise linear curve was fit to the measured power for each wind speed to approximate the relationship between wind speed and ice-free (clean) power (PWRclean; see Fig. 7a). During the winter season, the air-density effect is likely to increase the power production compared to autumn, as seen by the upward shift in maximum power production at moderate wind speeds in Fig. 7b. Correction for air-density will be considered in future LOWICE versions. Power curves like the one in Fig. 7a are applied to LAPS wind speeds to estimate PWRclean at each point in time. Though results for only one site are presented in this paper, unique power curves were developed for several wind farms across Sweden, where testing was performed. As described earlier, it is well known that icing can have a dramatic effect on power production, but quantification of these effects has proven to be difficult (e.g., Durstewitz et al. 2008). One simple approach is to assume that as soon as any amount of ice is expected on the turbine, that turbine will not produce power. This is overly simplistic, since power can still be produced when ice accretions are present, though the production rate is often reduced. With this in mind, a simple first version of power loss code was developed for LOWICE (version V0), according to following model. If no ice load is present, then no loss of power is expected. As the ice load increases from 0.0 to 10.0 kg m−1, the power loss increases linearly from 0% to 100% [i.e., “iced power production” (PWRiced) decreases linearly from 100% to 0%]. Once the ice load reaches 10.0 kg m−1, then a 100% power loss is assumed and PWRiced goes to zero. This overly simple approach allows the turbine to produce power at a reduced rate until a substantial ice load is expected,

 
formula

However, intense scrutiny of real-time observations from a wide variety of icing events at wind farms in Sweden made it clear that icing-induced power losses were poorly correlated with both measured and predicted ice loads. This was true when examining all data across several icing seasons and for numerous subperiods, including periods where the measured ice loads seemed to be of particularly good quality (i.e., loads were present and ice was clearly evident in webcam images; e.g., Fig. 8). Scatterplots and statistical analysis comparing ice loads and power losses for long periods and case studies showed poor results. Only weak trends were found and r-squared values only reached 0.12 at best. It was also discovered that significant, if not full, power could be produced when significant measured ice loads were present. Further, it was quite obvious that tying power loss to ice load during periods of persistent load sometimes resulted in gross overestimates of power loss and underestimates of power production. In a climate like Scandinavia’s, such loads can sometimes persist for weeks and even months at a time. Therefore, it was clearly evident that to better estimate icing-induced power losses, the losses have to be related to something other than ice load alone.

Fig. 7.

Power curves as percent of full power (y axis) against wind speed (x axis). (a) Ice-free turbine power plotted against wind speed during an ice-free sample period in 2011 (gray dots), and the linear fit generated by wind turbine measurements of wind speed and power production (black line). (b) As in (a), but for a period with icing during 2011.

Fig. 7.

Power curves as percent of full power (y axis) against wind speed (x axis). (a) Ice-free turbine power plotted against wind speed during an ice-free sample period in 2011 (gray dots), and the linear fit generated by wind turbine measurements of wind speed and power production (black line). (b) As in (a), but for a period with icing during 2011.

Fig. 8.

Webcam images taken on (left) 7 and (right) 8 Jan 2013 from the top of the turbine nacelle, illustrating the visual assessment of icing and showing the ice load instrument with a cone of ice on it (middle of both images).

Fig. 8.

Webcam images taken on (left) 7 and (right) 8 Jan 2013 from the top of the turbine nacelle, illustrating the visual assessment of icing and showing the ice load instrument with a cone of ice on it (middle of both images).

Further examination of observations indicated that other factors appeared to correlate more readily with power loss. In particular, power losses tended to be most evident when icing was active (i.e., ice was actively accreting at the site) and the intensity of the power losses was more strongly related to the icing rate than icing load. When icing is active, the ice is growing on the turbine blade, causing lift and drag penalties. As icing becomes inactive, the ice is often sublimated, melted, or shed from the blades fairly quickly as they rotate, resulting in a rapid return to power. The exception tends to be when turbine rotation has slowed dramatically or stopped altogether, sharply decreasing the chance for centrifugal shedding. Because of a relative lack of centrifugal force, ice on the slowly rotating vertical cylinders (used to measure loads) tends to remain in place, even when turbines shed ice and power returns. Thus, a much better correlation was found between power losses and icing rate than with icing load on a cylinder.

Time is also a critical element. First, as long as icing is active, its effect on power tends to increase or at least be maintained. Second, the longer the period of active icing is, the larger the effects tend to be. In LOWICE, this is called the ice “building effect.” The building effect is essentially a function of icing rate, time, and temperature. Analysis of observations (presence/absence of active icing, ice load, temperature, relative humidity, wind speed, ceiling height, visibility, webcam images, and model output) during and after icing events clearly indicated that once icing becomes inactive, depletion effects, such as sublimation and melting, begin to take hold. The stronger those effects are and the longer they persist, the more rapidly the power tends to return. In LOWICE, this is called the ice “clearing effect.” The clearing effect is a combination of the melting and sublimation effects, both of which are tied to their duration and intensity (sublimation rate and melting rate), plus the more stochastic and relatively instantaneous effect of shedding.

The equations provided below are empirical in nature and are derived from the examination of dozens of periods of active and inactive icing. They are meant to capture the essence of the effects of ice growth, decay, and their rates as associated with the presence and absence of supercooled liquid water, temperature, relative humidity, and wind speed present when events and nonevents are occurring, though it will be demonstrated in this article that these approximations of the physical process that drive icing and its effects perform reasonably well when judged against independent observations. The equations below can most certainly be refined based on the scientific literature.

 
formula
 
formula
 
formula
 
formula
 
formula
 
formula
 
formula
 
formula
 
formula
 
formula
 
formula
 
formula
 
formula
 
formula
 
formula
 
formula
 
formula

With these concepts in mind, LOWICE’s alternative power loss schemes (V1 and V2) combine the building and clearing effects described above [Eqs. (7), (8)]. At each time step, these factors are combined to determine the fractional expected loss factor (LOSSfactor; V1Loss or V2Loss, respectively), on a 0.0 (no power loss expected) to 1.0 (total power loss expected) scale. Iced power (PWRiced) is then calculated accordingly:

 
formula

Using this approach, icing-related losses can rapidly deplete power during periods of active icing, depending on the icing rate (i.e., intensity) and longevity of the event. Also, once active icing ceases, power can rapidly recover, depending on the rate of melting, sublimation, shedding, and the duration of icing inactivity.

3. Results and verification

As part of a wind pilot program, observations of power production and other fields (e.g., wind speed, temperature) have been examined for several wind farms across Sweden over five icing seasons: 2009–14. As a demonstration, the results are shown for one wind farm and for a single month (January 2013). This period has been chosen because it demonstrates numerous interesting features when observations were available continuously, including webcam images, turbine-mounted instruments, and power measurements. Because of the confidentiality agreements with wind power companies, the actual production, wind farm names, and certain other data could not be included in this paper. The wind farm turbines were located at various elevations around 800 m, while the corresponding LOWICE model height (i.e., representative hub height) was about 10 m higher for this specific wind farm. There were 712 h of matched data during January 2013. Temperature and wind speed verification have been made against measurements from instruments mounted at the top of one of the wind turbines. Manual, hourly assessments of the presence of ice and the occurrence of active icing were determined by independent visual inspection of webcam images in combination with inspection of measurements of visibility, ceiling height, and the other fields described above. Though the presence of ice and active icing can be particularly difficult in some environments (e.g., when a thin glaze of icing may be present or lighting is poor), the manual assessment dataset is generally quite reliable. Results of the ice detection and estimated power are described in section 3b. In section 3c, a statistical evaluation of icing detection and power losses is presented.

a. Temperature and wind

Analyzed and observed temperature fields were in fairly good agreement during the test period, though there was a persistent cold bias, averaging 2.4°C (Table 1; Fig. 9a). This bias appears to have been driven by a cold bias in ECMWF surface temperature forecasts (i.e., first-guess field for LAPS), which were found to be about 2°C in the −13° to +5°C range for mountain stations (at height > 300 m above mean sea level; Fig. 10a). Even though the LAPS analysis generally performs very well, there are still outliers with stations being less corrected, giving a slight cold bias in the −13° to +5°C temperature range (Fig. 10b). Modeled and observed winds were also quite comparable, with only a slight positive wind bias of 1.1 m s−1 (Table 1; Fig. 9b). At times, modeled winds were substantially higher than observed winds (e.g., 1–2, 17–19, and 24–26 January), but observed winds were sometimes falsely low due to icing effects on the anemometer (when the light green line is well below the gray lines in Fig. 9b). These false low “truth” values for wind contributed to the apparent 1.1 m s−1 high wind speed bias.

Table 1.

Statistics of the average values from the wind farm (a total of 18 turbines) during January 2013, where T is temperature (°C), U is wind speed (m s−1), and Pwr loss is power production losses (%) due to icing on turbine. Estimated power (Pwr) loss is given for LOWICE versions V0, V1, and V2 in the table.

Statistics of the average values from the wind farm (a total of 18 turbines) during January 2013, where T is temperature (°C), U is wind speed (m s−1), and Pwr loss is power production losses (%) due to icing on turbine. Estimated power (Pwr) loss is given for LOWICE versions V0, V1, and V2 in the table.
Statistics of the average values from the wind farm (a total of 18 turbines) during January 2013, where T is temperature (°C), U is wind speed (m s−1), and Pwr loss is power production losses (%) due to icing on turbine. Estimated power (Pwr) loss is given for LOWICE versions V0, V1, and V2 in the table.
Fig. 9.

Time series for January 2013 of (a) temperature and (b) winds speed from LOWICE [T-LAPS (red line) and WSP-LAPS (green line)]; observations of temperature and wind speed from standard instrumentation [T-OBS (tan line) and WSP-OBS (light green line)]; and individual turbine measurements, temperature and wind speed (T-01–T-09 and WS-01–WS-09) are shown with gray lines. All data are from the approximately same height (i.e., height of wind farm).

Fig. 9.

Time series for January 2013 of (a) temperature and (b) winds speed from LOWICE [T-LAPS (red line) and WSP-LAPS (green line)]; observations of temperature and wind speed from standard instrumentation [T-OBS (tan line) and WSP-OBS (light green line)]; and individual turbine measurements, temperature and wind speed (T-01–T-09 and WS-01–WS-09) are shown with gray lines. All data are from the approximately same height (i.e., height of wind farm).

Fig. 10.

Scatterplot of mountain stations’ (i.e., >300 m MSL) 2-m temperature for January 2013. (a) ECMWF on y axis and observations on x axis and (b) LAPS analysis on y axis and observed values on x axis. Colored dots represent the number of observations. Black dotted line is the mean value of dataset. Included is the 1:1 best-fit curve (thin black curve).

Fig. 10.

Scatterplot of mountain stations’ (i.e., >300 m MSL) 2-m temperature for January 2013. (a) ECMWF on y axis and observations on x axis and (b) LAPS analysis on y axis and observed values on x axis. Colored dots represent the number of observations. Black dotted line is the mean value of dataset. Included is the 1:1 best-fit curve (thin black curve).

b. Icing and power loss

The estimated observed power loss for January 2013 was 19.5% for all turbines at the site combined. Actual individual turbine power losses ranged from 3% to 34%. LOWICE expected power losses were quite reasonable for V1 (18.3%) and V2 (15.9%), but losses were grossly overestimated by the load-driven V0 (74.5%). Time series data shown in Fig. 11 reveal that V1 and V2 power losses tended to be quite similar and followed the trends of turbine observations quite well, while V0 frequently overestimated losses (e.g., 100% loss between 1 and 16 January), including when its iced power trends tracked somewhat similarly to V1 and V2 (25–31 January).

Fig. 11.

Time series for January 2013 of the percentage of maximum power from wind turbines (gray), LOWICE clean power (blue), and iced power (red, brown, and pink lines for loss code versions V2, V1, and V0, respectively). These traces are not visible when power values equal zero. Color bars at the bottom provide an hour-by-hour assessment of the presence of icing as follows. Bar 1: LOWICE-expected icing [red represents active icing (rate ≥ 10 g h−1), yellow denotes inactive icing (ice present but rate <10 g h−1), green denotes no ice, and white denotes no information]. Bar 2: manually assessed icing through webcam (red represents active icing, yellow denotes active icing possible but not confirmed, green denotes no active icing, and white denotes no information).

Fig. 11.

Time series for January 2013 of the percentage of maximum power from wind turbines (gray), LOWICE clean power (blue), and iced power (red, brown, and pink lines for loss code versions V2, V1, and V0, respectively). These traces are not visible when power values equal zero. Color bars at the bottom provide an hour-by-hour assessment of the presence of icing as follows. Bar 1: LOWICE-expected icing [red represents active icing (rate ≥ 10 g h−1), yellow denotes inactive icing (ice present but rate <10 g h−1), green denotes no ice, and white denotes no information]. Bar 2: manually assessed icing through webcam (red represents active icing, yellow denotes active icing possible but not confirmed, green denotes no active icing, and white denotes no information).

LOWICE estimates of active icing (ICErate ≥ 10 g h−1) and inactive icing (ice present but ICErate < 10 g h−1) are indicated on the color bars near the bottom of Fig. 11. The periods of active and inactive icing were generally captured very well, though active icing was indicated more often by LOWICE than by manual assessments. Most of these overwarnings were for short periods; however, there were a few prolonged periods when LOWICE indicated icing but none was observed (e.g., 1–2, 18–19, and 24–26 January in Fig. 11; compare red portions of the two color bars). The second of these expected events resulted in significant expected power losses. The power time series chart implies that LOWICE’s iced power forecasts were still reasonable, but this was caused by a combination of overestimated wind speeds (see Fig. 9b) and overestimated icing effects compensating for one another. This is an example of how iced wind power models can get the right answer for the wrong reason and why statistical comparisons alone are not enough to determine the quality of system output. Despite this example, overall power losses for icing-rate-driven V1 and V2 were quite comparable to observed losses and far superior to those from the ice-load-driven V0 power loss code.

c. Statistical results

Though statistical results have their shortcomings, it is important to attempt to measure the quality of the icing and power fields produced by LOWICE and, indeed, any such system. To this end, standard statistical measures were used to determine the probability of detecting both positive (“Yes”) and negative (“No”) icing events and power loss events (PODy, PODn, respectively, following Mahoney et al. 2002; see Table 2). Both measures and their derivatives [e.g., critical success index (CSI) and true skill score (TSS)] are important because an icing and power assessment system must capture both the presence and absence of icing and its downstream effect on power to be useful to those making decisions regarding the operation of turbines in cold climates and the trading of power expected to be available to the grid.

Table 2.

Statistical parameters and explanations thereof (Mahoney et al. 2002). PODy and PODn measure the percentage of times that the system correctly forecast the occurrence of positive (Yes) and negative (no) events. It is best to maximize the PODy and PODn values. FAR measures the frequency of Yes forecasts that did not verify. FAR should be minimized, because false alarms indicate overforecasting of positive (Yes) events. CSI measures the number of correct Yes forecasts relative to the total number of Yes forecasts and observations. CSI value should be maximized. TSS summarizes the ability of the forecasts, objectively measuring their skill, because it combines the PODy and PODn measures. Models are not rewarded that have high PODn by virtue of underforecasting events (and thus have a low PODy) or high PODy by virtue of overforecasting (and thus have a low PODn). Positive TSS value indicates skill, zero indicates no skill, and negative value indicates negative skill.

Statistical parameters and explanations thereof (Mahoney et al. 2002). PODy and PODn measure the percentage of times that the system correctly forecast the occurrence of positive (Yes) and negative (no) events. It is best to maximize the PODy and PODn values. FAR measures the frequency of Yes forecasts that did not verify. FAR should be minimized, because false alarms indicate overforecasting of positive (Yes) events. CSI measures the number of correct Yes forecasts relative to the total number of Yes forecasts and observations. CSI value should be maximized. TSS summarizes the ability of the forecasts, objectively measuring their skill, because it combines the PODy and PODn measures. Models are not rewarded that have high PODn by virtue of underforecasting events (and thus have a low PODy) or high PODy by virtue of overforecasting (and thus have a low PODn). Positive TSS value indicates skill, zero indicates no skill, and negative value indicates negative skill.
Statistical parameters and explanations thereof (Mahoney et al. 2002). PODy and PODn measure the percentage of times that the system correctly forecast the occurrence of positive (Yes) and negative (no) events. It is best to maximize the PODy and PODn values. FAR measures the frequency of Yes forecasts that did not verify. FAR should be minimized, because false alarms indicate overforecasting of positive (Yes) events. CSI measures the number of correct Yes forecasts relative to the total number of Yes forecasts and observations. CSI value should be maximized. TSS summarizes the ability of the forecasts, objectively measuring their skill, because it combines the PODy and PODn measures. Models are not rewarded that have high PODn by virtue of underforecasting events (and thus have a low PODy) or high PODy by virtue of overforecasting (and thus have a low PODn). Positive TSS value indicates skill, zero indicates no skill, and negative value indicates negative skill.

1) Presence of ice and active icing

As described above, the presence and absence of ice, as well as whether the ice was actively growing, were assessed by subjective, manual inspection of a combination of webcam images and time history of automated measurements of temperature, dewpoint, visibility, ceiling height, and ice load. In addition, objective assessments were made using measured icing-related parameters and derivatives thereof. Both approaches were used to produce Yes/No “ground truth” data on the presence of ice and active icing at the wind farm. These include increases in observed ice load over time (1-, 3-h periods) and the simultaneous presence of T < 0°C and either visibility < 1000 m or ceiling height < 250 m (see Table 3). An absolute Yes/No threshold that indicates that icing has occurred does not exist, so a series of thresholds is applied to reach the final answer. The LOWICE system uses similar parameters to determine the presence of ice (see Table 4; Fig. 12). Normally, a threshold of 0.1 N would be used for the observed presence of ice on the reference cylinder weighing gauge, but the presence of a bias in the observed loads at this site necessitated the use of a higher threshold (4.5 N).

Table 3.

Observed parameters and thresholds used for objective assessment of the presence of ice and active icing. These include measured icing load (N), hourly and 3-hourly changes in icing load, the simultaneous presence of temperatures that were subfreezing or nearly so (T-factor) with either visibility less than 1000 m or ceiling height less than 250 m (vis-factor).

Observed parameters and thresholds used for objective assessment of the presence of ice and active icing. These include measured icing load (N), hourly and 3-hourly changes in icing load, the simultaneous presence of temperatures that were subfreezing or nearly so (T-factor) with either visibility less than 1000 m or ceiling height less than 250 m (vis-factor).
Observed parameters and thresholds used for objective assessment of the presence of ice and active icing. These include measured icing load (N), hourly and 3-hourly changes in icing load, the simultaneous presence of temperatures that were subfreezing or nearly so (T-factor) with either visibility less than 1000 m or ceiling height less than 250 m (vis-factor).
Table 4.

LOWICE output used for determining the presence of ice and active icing.

LOWICE output used for determining the presence of ice and active icing.
LOWICE output used for determining the presence of ice and active icing.
Fig. 12.

Statistics for (a)–(c) icing and (d)–(f) power loss (all turbines) at the site for January 2013. For icing, PODy, PODn, FAR, CSI, and TSS were calculated by comparing LOWICE indications of the presence or absence of any ice (Yes when expected load was nonzero) and active icing (Yes when ICErate > 10 g h−1) to observations, including (a) manual assessments (visual inspection) of the presence of ice, ice growth (frame-by-frame examination of webcam images), and measurements of (b) dLoad/dt (hourly; Yes when > 0 N h−1), dLoad/dt (3 hourly; Yes when >0 N 3 h−1), and “T-vis” (normally observed T and visibility, but T and ceiling height were used at this site because visibility was not measured there; Yes when −40°C < T < +0.125°C and ceiling height < 125 m). For power loss, the same statistical fields were calculated by comparing the LAPS–LOWICE “expected” losses and the “observed” losses based on turbine measurements of wind speed and power production, and then applying 10%, 30%, 50%, 70%, and 90% power loss thresholds. A marker is plotted for each combination (e.g., 10% expected and 10% observed loss, 10% expected and 30% observed, etc.). Observed losses were estimated by comparing the observed power to the “expected clean power,” which was calculated by passing the observed wind speed through the power curve.

Fig. 12.

Statistics for (a)–(c) icing and (d)–(f) power loss (all turbines) at the site for January 2013. For icing, PODy, PODn, FAR, CSI, and TSS were calculated by comparing LOWICE indications of the presence or absence of any ice (Yes when expected load was nonzero) and active icing (Yes when ICErate > 10 g h−1) to observations, including (a) manual assessments (visual inspection) of the presence of ice, ice growth (frame-by-frame examination of webcam images), and measurements of (b) dLoad/dt (hourly; Yes when > 0 N h−1), dLoad/dt (3 hourly; Yes when >0 N 3 h−1), and “T-vis” (normally observed T and visibility, but T and ceiling height were used at this site because visibility was not measured there; Yes when −40°C < T < +0.125°C and ceiling height < 125 m). For power loss, the same statistical fields were calculated by comparing the LAPS–LOWICE “expected” losses and the “observed” losses based on turbine measurements of wind speed and power production, and then applying 10%, 30%, 50%, 70%, and 90% power loss thresholds. A marker is plotted for each combination (e.g., 10% expected and 10% observed loss, 10% expected and 30% observed, etc.). Observed losses were estimated by comparing the observed power to the “expected clean power,” which was calculated by passing the observed wind speed through the power curve.

For active icing, PODy values were all between 0.55 and 0.81, PODn values were all near 0.55, CSIs were 0.18–0.35, and TSSs were 0.11–0.36 (Figs. 12a–c). Overall, these results show some (but not particularly impressive) skill for this particular test site and period. However, it is important to note that the LOWICE system’s results were best for comparisons with the most reliable observed field: manually assessed active icing (PODy: 0.81, PODn: 0.56, CSI: 0.25, TSS: 0.36). However, FAR was high at 0.73. For the presence of ice, LOWICE indicated positive loads throughout the month of January 2013, so PODy was 1.0, while PODn was 0.0, resulting in the point in the bottom-right portion of Fig. 12a. It is interesting to note that although the manual assessment indicated some hours with no ice presence, the icing load was measured to be >4.5 N throughout the month, so PODn and TSS were undefined for this field (which gives one less verification point in Figs. 12a,c). The CSI for manually assessed ice presence was 0.89.

2) Statistical results—Power loss

LOWICE’s expected power loss estimates were verified against estimates of observed power loss. Again, there is no absolute Yes/No threshold that indicates that power loss has occurred or that it can be attributed to icing. However, a reasonable assessment of power loss can be gained by applying some simple thresholds. First, turbine-measured wind speeds must considered to be strong enough to be able to produce clean power that is at least 25% of the rated output from the turbine. This is estimated by sending the turbine-measured wind speed through LOWICE’s piecewise linear approximation of the wind farm’s power curve. The 25% criterion is used to be certain that an adequate amount of power is expected to allow for a meaningful estimate of power loss, as well as to avoid the inaccuracies of the piecewise linear power curve near the cut-in wind speed. Next, if at least 75% of the turbines met the clean power criterion mentioned above, then the measured power from each turbine was divided by the expected clean power to determine the percentage of expected clean power that is being produced. This value was subtracted from 1.0 to determine the power loss and then expressed as a percentage (see Table 5). Power loss thresholds (10%, 30%, 50%, 70%, and 90%) were then applied to the estimated losses described above. Each of these thresholds is used to determine when observed power loss was considered to be present (a Yes observation).

Table 5.

Power loss assessments for LOWICE system output to produce a Yes answer.

Power loss assessments for LOWICE system output to produce a Yes answer.
Power loss assessments for LOWICE system output to produce a Yes answer.

Among the turbines with adequate wind speeds to meet the 25% rated power output threshold described above, the number of turbines that met the power loss percentage threshold was counted. At least half of the turbines with adequate expected clean power had to meet the power loss threshold for a Yes observation of power loss to be indicated for that time. As stated above, for a Yes observation of power loss, multiple power loss thresholds (10%, 30%, …, 90%) were used and at least 75% of the turbines had to meet the 25% expected power threshold.

For the LOWICE system output to produce a Yes, the system’s expected iced power was compared to its expected clean power. Expected power loss was calculated by dividing the expected iced power by the expected clean power, and then the result was subtracted from 1.0 and expressed as a percentage (Table 5). Like observed power, power loss thresholds of 10%, 30%, 50%, 70%, and 90% were applied to convert expected losses from the system to a Yes/No answer. Yes and No forecasts and observations were compared for every combination of forecast and observed power loss threshold. For example, LOWICE expected losses of ≥10% were compared to observed losses of ≥0%, 30%, 50%, 70%, and 90%. The same was true for LOWICE expected losses of ≥30%, 50%, 70%, and 90%. This resulted in a matrix of Yes and No forecasts and observations of power losses that were used to generate the plots in Figs. 12d–f.

Power loss results were generally quite good, especially for V1 and V2. Results are threshold dependent, but for V1 and V2 the PODn values were >0.65 in all cases, PODy values for most thresholds exceeded 0.3, and some exceeded 0.8. CSI and TSS values suggest that the system had good skill for many power loss thresholds. The load-based V0 system performed more poorly, with generally higher PODy (due to the persistent presence of positive loads) at the expense of lower PODn, higher FAR, and generally lower CSI and TSS (Figs. 12d–f).

4. Additional applications

Beyond the real-time power and icing assessments described in this paper, the LAPS–LOWICE system may also prove useful as part of the assessment of planned wind power sites and as an ice detection method.

Both wind and icing evaluations should be done as part of the evaluation of planned wind parks in order to assess their potential for power production and both the need for and potential benefits of an anti-icing or deicing system. Different ice detection measurement approaches give different information on frequency and duration of icing events, with none proving to be superior to the other (Tammelin et al. 2005). Hence, applying multiple methods would theoretically improve the reliability of icing information. Long-term estimations of the frequency of icing from a system such as LAPS–LOWICE could serve as one method for assessing the expected wind power and icing at a new site. In fact, LAPS–LOWICE has been used to generate 5-yr climatologies for multiple sites across Sweden using a combination of archived GFS model output and observations from satellite and surface stations. Because of their proprietary nature, the results of these studies cannot be included here.

In cold climate regions (i.e., where icing occurs) many turbines have heating systems, which need a control strategy. Proper and fast identification of icing events in real time is crucial to operating these systems properly and maximizing power production, since power production losses of 5%–15% can occur early in icing events before deicing systems are activated (Peltola et al. 1996). Usually, a basic control strategy includes an ice detection method that is used to activate the system. Detection strategies may rely upon 1) the difference between the wind speeds measured by heated and unheated anemometers that exceed a certain limit and 2) measured power that is lower than expected ice-free power for a given wind speed. Different ice detection methods give different results, and no method is accurate and reliable for all situations (Marjaniemi et al. 2000). Therefore, usage of several different ice detection methods may be beneficial for the control of wind turbines. LOWICE may prove to be useful for this purpose.

5. Discussions, conclusions, and future work

In this paper new methods on how to estimate wind power production and the effects of icing thereon have been described. Using LAPS 3D analyses of temperature, winds, and clouds, the LOWICE system has proven to be able to estimate expected power from wind turbines, power losses that are associated with icing, and the recovery of power associated with the depletion of icing effects. Examination of data from the active icing month of January 2013 showed that LAPS–LOWICE provided realistic results for the fields described above. However, this paper does not focus only on good results to demonstrate the robustness of the system. Instead, it describes both good and poor results and some of the root causes of them, in an effort to move the science of wind power prediction forward. Examples include the anomalously cold LAPS temperature bias and both the over- and underdiagnosis of icing and their effects on expected production.

ECMWF surface temperature forecasts were a potential cause of the overall negative temperature bias in LAPS output (see section 3a; Fig. 10). This may have been particularly important for power loss errors early in the month, when observed temperatures hovered around 0°C, yet icing loads persisted in the system. Another potential source of the cold bias may be due to the assimilation of surface temperatures from nearby mountain valley stations, with locally very cold temperatures (i.e., valley inversion effects), which might not be representative of the local climate at an elevated wind farm. Also, it is possible that wind turbine temperature sensors may have been affected by heating from deicing equipment, which might have affected the validation results. At several sites, turbine-measured temperatures proved to be consistently warmer than independently measured temperatures at essentially the same height (Fig. 9a). Wind speed verification results were quite good, and both the timeliness and amplitude of important wind features were captured quite well by LAPS in time series analysis. The largest exceptions appeared to be associated with falsely low wind observations from iced anemometers. Uncertainties in measured temperatures and winds will be further investigated and reported in upcoming articles.

Verification of power estimates and icing effects on power indicated that V1 and V2 of the power loss code performed particularly well, providing consistent, reasonable estimates of power production. In contrast, V0 performed relatively poorly because its power losses are tied solely to the expected ice load, which was overestimated. The ice load has also proven to be poorly correlated with power loss.

Future developments of the LAPS–LOWICE system will include 1) the use of radar data for better cloud and precipitation analyses and SLWC adjustment and 2) an estimation of insolation to better assess the potential for ice shedding. Recent measurements at wind farms could be used to correct the LAPS–LOWICE temperatures, wind speeds, and clean power estimates. Also, observations of visibility, ceiling height, and icing (e.g., from Holooptics, ice-load trends, or ice detectors such as the “Rosemount” probe; e.g., Mughal and Virk 2013) could be used to improve system estimates of the presence and intensity of icing conditions at the site.

While real-time diagnoses of the wind speed, icing, and power from LOWICE clearly have utility, there is also great value in accurate and timely predictions of these parameters. Thus, a forecast version of LOWICE, known as “FLOWICE,” has been developed using numerical model output to generate forecasts of wind speed, icing, and wind power production out to 48 h. In an effort to improve initialization of icing and power loss fields at the start of each run, FLOWICE runs are initiated using the most recent LOWICE analysis fields, rather than “cold start” models that are initialized with no icing or model runs that are initiated from the final state of the previous model run, which can lead to compounding effects of previous poor forecasts. Corrections to system initialization and biases could also be gained from comparison to recent and historical observations. This concept was tested during the 2014–15 icing season and these adjustments appear to be helpful.

The analysis and forecasting of icing is very difficult, especially close to ground. There is a great value in real-time observations if they are applied correctly and blended effectively with model fields. In particular, all data (observations and model fields) related to clouds and precipitation (satellite, METAR, SYNOP, T and RH profiles, etc.) have proven to be of great interest and can have great value for wind turbine icing and power production diagnoses and forecasts. The developing LAPS–LOWICE and FLOWICE systems have great potential to provide highly valuable information to the wind power community.

Acknowledgments

We want to thank OX2, the Swedish Energy Agency (funding), Göran Ronsten, the Wind Pilot Program Reference Group, and the anonymous reviewers of this paper for their valuable input and/or support of this work.

REFERENCES

REFERENCES
Albers
,
S. C.
,
J. A.
McGinley
,
D. L.
Birkenheuer
, and
J. R.
Smart
,
1996
:
The Local Analysis and Prediction System (LAPS): Analyses of clouds, precipitation, and temperature
.
Wea. Forecasting
,
11
,
273
287
, doi:.
Bernstein
,
B. C.
,
F.
McDonough
,
M. K.
Politovich
,
B. G.
Brown
,
T. P.
Ratvasky
,
D. R.
Miller
,
C. A.
Wolff
, and
G.
Cunning
,
2005
:
Current icing potential (CIP): Algorithm description and comparison with aircraft observations
.
J. Appl. Meteor.
,
44
,
969
986
, doi:.
Bernstein
,
B. C.
,
F.
McDonough
,
C. A.
Wolff
,
M. K.
Politovich
,
G.
Cunning
,
S.
Mueller
, and
S.
Zednik
,
2006
: The new CIP icing severity product. 12th Conf. on Aviation, Range and Aerospace Meteorology, Atlanta, GA, Amer. Meteor. Soc., P9.5 [Available online at http://ams.confex.com/ams/pdfpapers/102273.pdf.]
Bernstein
,
B. C.
,
E.
Gregow
,
I.
Wittmeyer
, and
J.
Hirvonen
,
2011a
: LOWICE: An example of the adaption of in-flight icing diagnosis concepts to the structural icing environment. Proc. SAE 2011 Int. Conf. on Aircraft and Engine Icing and Ground Deicing, Chicago, IL, SAI International and AIAA, 2011-38-0070. [Available online at http://papers.sae.org/2011-38-0070/.]
Bernstein
,
B. C.
,
C. A.
Wolff
, and
F. M.
McDonough
,
2011b
: A regional comparison of icing conditions in boundary layer clouds. Proc. 2011 SAE Int. Conf. on Aircraft and Engine Icing and Ground Deicing, Chicago, IL, SAI International and AIAA, 2011-38-0021. [Available online at http://papers.sae.org/2011-38-0021/.]
Cattin
,
R.
,
S.
Kunz
,
A.
Heimo
,
G.
Russi
,
M.
Russi
, and
M.
Tiefgraber
,
2007
: Wind turbine ice throw studies in the Swiss Alps. Proc. European Wind Energy Conf. and Exhibition 2007, Milan, Italy, EWEA, BL3.269. [Available online at http://www.meteotest.ch/fileadmin/user_upload/Windenergie/pdfs/paper_ewec2007_cattin_final.pdf.]
Derrien
,
M.
, and
H.
Le Gléau
,
2005
:
MSG/SEVIRI cloud mask and type from SAFNWC
.
Int. J. Remote Sens.
,
26
,
4707
4732
, doi:.
Durstewitz
,
M.
,
J.
Dobschinski
, and
Z.
Khadiri-Yazami
,
2008
: Wind power forecast accuracy under icing conditions—General approach, practical applications and options for considering effects of wind turbine icing. Proc. Winterwind 2008, Norrköping, Sweden, Swedish Wind Power Association [Available online at http://www.winterwind.se/2008/presentationer/17_Durstewitz_Winterwind_2008.pdf.]
Dybbroe
,
A.
,
K.-G.
Karlsson
, and
A.
Thoss
,
2005
:
NWCSAF AVHRR cloud detection and analysis using dynamic thresholds and radiative transfer modeling. Part II: Tuning and validation
.
J. Appl. Meteor.
,
44
,
55
71
, doi:.
ECMWF
,
2014
: IFS documentation CY40r1. Accessed 8 May 2014. [Available online at http://old.ecmwf.int/research/ifsdocs/CY40r1/>.]
EWEA
,
2013
: Wind power growth expected to slow in 2013, but recovery predicted. Accessed 6 April 2014. [Available online at http://www.ewea.org/blog/2013/04/wind-power-growth-expected-to-slow-in-2013-but-recovery-predicted/.]
Finstad
,
K. F.
,
E. P.
Lozowski
, and
E. M.
Gates
,
1988
:
A computational investigation of water droplet trajectories
.
J. Atmos. Oceanic Technol.
,
5
,
160
170
, doi:.
GWEC
,
2012
: Global wind report 2012: Annual market update. 72 pp. [Available online at http://www.gwec.net/wp-content/uploads/2012/06/Annual_report_2012_LowRes.pdf.]
Haggerty
,
J. A.
,
F.
McDonough
,
J.
Black
,
S.
Landolt
,
C.
Wolff
,
S.
Mueller
,
P.
Minnis
, and
W. L.
Smith
,
2008
: Integration of satellite-derived cloud phase, cloud top height, and liquid water path into an operational aircraft icing nowcasting system. 13th Conf. on Aviation, Range and Aerospace Meteorology, New Orleans, LA, Amer. Meteor. Soc., 3.13. [Available online at https://ams.confex.com/ams/pdfpapers/131893.pdf.]
Huffman
,
G. J.
, and
G. A.
Norman
Jr.
,
1988
:
The supercooled warm rain process and the specification of freezing precipitation
.
Mon. Wea. Rev.
,
116
,
2172
2182
, doi:.
ISO
,
2001
: Atmospheric icing of structures. International Organization for Standarization Copyright Office ISO 12494:2001, 56 pp. [Available online at http://www.iso.org/iso/catalogue_detail.htm?csnumber=32823.]
Koskinen
,
J. T.
, and Coauthors
,
2011
:
The Helsinki Testbed: A mesoscale measurement, research, and service platform
.
Bull. Amer. Meteor. Soc.
,
92
,
325
342
, doi:.
Le Bot
,
C.
,
2004
: SIGMA: System of Icing Geographic Identification in Meteorology for Aviation. 11th Conf. on Aviation, Range and Aerospace Meteorology, Hyannis, MA, Amer. Meteor. Soc., P6.5. [Available online at https://ams.confex.com/ams/11aram22sls/techprogram/paper_81704.htm.]
Mahoney
,
J.
,
J. K.
Henderson
,
B. G.
Brown
,
J. E.
Hart
,
A.
Loughe
,
C.
Fischer
, and
B.
Sigren
,
2002
: The Real-Time Verification System (RTVS) and its application to aviation weather forecasting. Preprints, 10th Conf. on Aviation, Range and Aerospace Meteorology, Portland OR, Amer. Meteor. Soc., 9.8. [Available online at https://ams.confex.com/ams/13ac10av/techprogram/paper_40728.htm.]
Makkonen
,
L.
,
2000
:
Models for the growth of rime, glaze, icicles and wet snow on structures
.
Philos. Trans. Roy. Soc. London
,
358A
,
2913
2939
, doi:.
Marjaniemi
,
M.
,
H.
Holttinen
,
J.
Keinanen
,
E.
Holttinen
,
L.
Makkonen
,
E.
Peltola
,
T.
Maki
, and
K. O.
Petersen
,
2000
: Wind turbines in light icing conditions—Experiences of the Pori 8 MW wind farm. Proceedings of the BOREAS V Conference, B. Tammelin et al., Eds., Finnish Meteorological Institute, 13 pp.
Mughal
,
U. N.
, and
M. S.
Virk
,
2013
: Atmospheric icing sensors—An insight. SENSORCOMM 2013: The Seventh International Conference on Sensor Technologies and Applications, S. Yurish and M. S. Virk, Eds., IARIA, 191–199. [Available online at http://www.thinkmind.org/index.php?view=article&articleid=sensorcomm_2013_8_20_10053.]
Peltola
,
E.
,
M.
Marjaniemi
,
J.
Kaas
, and
E.
Aarnio
,
1996
: Pyhätunturi operational experiences. BOREAS III: Proceedings of an International Meeting, B. Tammelin et al., Eds., Finnish Meteorological Institute,
131
146
.
Politovich
,
M. K.
, and
B. C.
Bernstein
,
1995
:
Production and depletion of supercooled liquid water in a Colorado winter storm
.
J. Appl. Meteor.
,
34
,
2631
2648
, doi:.
Portin
,
H. J.
,
M.
Komppula
,
A. P.
Leskinen
,
S.
Romakkaniemi
,
A.
Laaksonen
, and
K. E. J.
Lehtinen
,
2009
:
Observations of aerosol–cloud interactions at the Puijo semi-urban measurement station
.
Boreal Env. Res.
,
14
,
641
653
.
Rogers
,
R. R.
, and
M. K.
Yau
,
1989
: A Short Course in Cloud Physics. 3rd ed. International Series in Natural Philosophy, Vol. 113, Butterworth Heinemann, 304 pp.
SAFNWC
,
2013
: Algorithm theoretical basis document for “cloud products” (CMa-PGE01 v3.2, CT-PGE02 v2.2 & CTTH-PGE03 v2.2). SAFNWC SAF/NWC/CDOP2/MFL/SCI/ATBD/01, Issue 3, Rev. 2.1, 87 pp. [Available online at http://www.nwcsaf.org/indexScientificDocumentation.html.]
Tafferner
,
A.
,
T.
Hauf
,
C.
Leifeld
,
T.
Hafner
,
H.
Leykauf
, and
U.
Voigt
,
2003
:
ADWICE: Advanced Diagnosis and Warning System for Aircraft Icing Environments
.
Wea. Forecasting
,
18
,
184
203
, doi:.
Tammelin
,
B.
,
M.
Cavaliere
,
H.
Holttinen
,
C.
Morgan
,
H.
Seifert
, and
K.
Säntti
,
2000
: Wind energy production in cold climate (WECO). Finnish Meteorological Institute Rep. JOR3-CT95-0014, 38 pp. [Available online at http://cordis.europa.eu/documents/documentlibrary/47698271EN6.pdf.]
Tammelin
,
B.
, and Coauthors
,
2005
: Wind turbines in icing environment: Improvement of tools for siting, certification and operation. Finnish meteorological Institute NEW ICETOOLS Rep. 2005:6, 127 pp.