The Weather Research and Forecasting model (WRF) is used to provide 6–12-h forecasts of the necessary input parameters to a separate algorithm that determines the most likely precipitation type at each model grid point. In instances where freezing rain is indicated, an ice accretion model allows forecasts of radial ice thickness to be developed. The resulting forecasts are evaluated for 38 icing events of varying magnitude that occurred in the eastern United States using National Weather Service storm impact reports and observed data from Automated Surface Observing Systems (ASOS). Ice accretion hindcasts, using the WRF, allow the development of climatologies based on archived model initialization data.
Ice accretion forecasts, based on the Ramer precipitation-type algorithm, consistently underestimated the maximum observed ice accretion amounts by between 10 and 20 mm. Ice accretion at ASOS sites was also underestimated. Applying a modification to the Ramer precipitation-type algorithm, and focusing on the thermal profile below the lowest 0°C isotherm, improved the ice accretion forecasts, but still underestimated the maximum ice thickness. Little bias was evident in ice accretion forecasts for the ASOS sites. Using previous observations from outside the forecast window to account for WRF and precipitation-type algorithm biases in precipitation amount, wind speed, temperature, and precipitation type provided some forecast improvement. The forecast procedure using the modified Ramer precipitation algorithm captures both the magnitude and extent of icing in both widespread severe icing events and localized storms. Minimal icing is indicated in events and at locations where precipitation fell as rain or snow.
The accretion of ice on components of the electric power transmission and distribution system is one of the primary impacts of freezing rain events. Electric power systems throughout the eastern United States have suffered repeated, often significant, damage from ice storms in recent years. In December 2002, for example, an ice storm caused more than 65% of Duke Power’s 2.2 million customers to lose power. In some instances these power outages lasted for more than 1 week. Recovery from this storm required 12 500 field and support personnel (R. Meffert 2003, personal communication). The January 1998 ice storm that affected locations throughout New York, northern New England, and southeastern Quebec resulted in economic losses that exceeded $1 billion (DeGaetano 2000). Nearly a quarter of these loses were associated with damage to electric utilities.
Electric power companies face many decisions in preparing and responding to major ice storms. Long before the approach of a storm, companies must determine how often and how much to trim trees along the length of the system, how many repair crews to keep on staff, and what quantity of repair material to keep in their inventory. As a storm approaches, companies are faced with deciding how many repair crews to mobilize, and whether to request crews from other companies. During (and immediately after) an ice storm, utilities must decide on how to prioritize repairs and estimate power restoration times.
These real-time decisions rely upon accurate ice accretion forecasts. Longer lead-time decisions require spatially interpolated climatological ice accretion fields at high resolution. The creation of such fields has typically relied on the statistical interpolation of station observations (e.g., Jones 1996). However, forecast model hindcasts, also have the potential to provide this information.
As a basis for these decisions, the engineering community has developed empirical models to predict the number of outages given an extreme weather event (e.g., Brown et al. 1997; Shen et al. 1999; Liu et al. 2008) or estimate restoration times following a weather-induced outage (Liu et al. 2007). Typically, these models have been applied to hurricane winds. In these cases, simulated hurricane wind fields are used to drive the model. Wind speeds at the spatial scale of zip codes are obtained using a combination of spatial data interpolation and the results of a geographic information system (GIS) based hurricane wind field simulation model (Huang et al. 2001). For a specific hurricane, the model was initialized with a preliminary wind field based on reconnaissance aircraft measurements, assumed surface-to-aircraft wind speed ratios, and a projected decay of the storm as it moved inland. The estimated winds were adjusted based on observations as the storm progressed.
Despite an interest by utilities in extending these models to ice-storm-related outages, an analogous means of producing ice accretion fields suitable for input into electric power system outage and restoration time models does not exist. Apparently, relatively little research has sought to develop techniques for forecasting ice accretions on land-based structures. Rather, the literature focuses on techniques to forecast precipitation type (e.g., Wandishin et al. 2005) and aircraft icing intensity (e.g., Tafferner et al. 2003) or estimate ice accretion based on meteorological observations (e.g., Jones et al. 2004). Nowcast techniques can make use of these observation-based ice accretion estimates (e.g., Thornes and Davis 2002). Statistical precipitation-type forecast techniques have also been proposed. Hux et al. (2001) use discriminant analysis to forecast mixed winter precipitation based on freezing level height and temperature. Larouche et al. (2000) use a neural network to “predict” ice accretion on overhead power lines. As their method relies on meteorological observations and real-time icing rate meter data, its forecast application is limited to nowcasts.
In this paper, we develop and test a hybrid approach for forecasting ice accretion on electric distribution lines. The method uses Jones’s (1996) ice accretion model as a basis, but substitutes hourly temperature, precipitation amount, and wind speed forecast values for the observations. In addition, three routines for forecasting precipitation type are evaluated, as the modified Jones (1996) freezing rain model is only applied during hours in which freezing rain is forecasted. The resulting ice accretion fields, when input into electric power restoration models, allow damage assessments to be made as the storm approaches, before such information is available from the field. Such forecasts allow companies to prioritize repairs to minimize per capita outage duration. The resulting forecasts also provide a means for better informing emergency management authorities and the public of expected poststorm restoration times. Hindcasted ice accretion fields can also be generated using the hybrid approach. When coupled with the electric distribution system models, such data, which are currently lacking, will enable utilities to assess the economic value of management practices such as tree trimming and distribution system maintenance.
2. Precipitation-type models
Cortinas and Baldwin (1999) describe and compare the performance of six precipitation-type forecast algorithms. Collectively, the methods were better able to forecast the occurrence of rain and snow than ice pellets or freezing rain. The methods tended to overforecast freezing rain occurrence as noted by the high probability of detection and high false alarm ratio. These results are supported by Cortinas et al. (2002), with the Ramer algorithm (Ramer 1993) being the least biased in terms of its freezing rain forecasts. Nonetheless, Wandishin et al. (2005) conclude that no precipitation-type algorithm is superior to the others under all conditions.
a. Ramer algorithm
Given its performance, in terms of freezing rain, as reported by Cortinas and Baldwin (1999), the Ramer algorithm is adopted here. Ramer (1993) provides an in-depth description of this algorithm, so only a brief overview is given. The algorithm is based on the ice fraction of precipitation as it reaches the ground. The highest saturated level (RH > 90%) in the model atmosphere is referred to as the precipitation generation level (PGL). The wet-bulb temperature (Tw) at the PGL determines the initial phase of precipitation. If Tw < −6.6°C, hydrometeors are considered to be entirely ice (ice fraction = 1); otherwise, liquid is assumed (ice fraction = 0). As a hydrometeor falls through each model atmosphere layer, melting or freezing occurs according to the equation
where I is the ice fraction, P is pressure (hPa); Tw is the layer-average wet-bulb temperature (°C), RH is the layer-average relative humidity, E′ is the Ramer empirically derived constant (=0.045°C), and n is the model layer.
Precipitation type is ultimately determined by the value of the ice fraction and by Tw at the lowest model level using empirical thresholds. Freezing rain is diagnosed if the hydrometeor is mostly liquid (I < 0.04) and Tw near the surface is less than 0°C. A freezing mix is assumed if 0.04 < I < 0.85 and Tw is less than 0°C.
b. Modified Ramer algorithm
The existence of freezing rain is most sensitive to atmospheric conditions near and at the surface. A second version of the precipitation-type algorithm focuses on the accumulated ice fraction below the lowest freezing layer. First, the original Ramer algorithm was executed to determine the most likely precipitation type. These estimates were then reevaluated by focusing only on the conditions near the surface. The lowest 0°C isotherm in the WRF model atmosphere was identified and changes in the ice fraction calculated at each model level between this layer and the surface to create a new variable termed the freezing level ice fraction, IFL. We set IFL to zero in the >0°C air above the surface subfreezing layer and it increases as the hydrometeor falls to the surface according to the formula
where K is the number of model atmosphere layers below the lowest 0°C isotherm.
If the original (unmodified) Ramer algorithm specifies freezing rain, then this precipitation type is retained. Freezing rain is also specified when 0.0 < IFL ≤ 0.1 and surface temperature Ts ≤ 0°C; otherwise, the unmodified Ramer precipitation type is specified. This may include a freezing mix. These empirical IFL thresholds were determined from multiple ice storm cases from New York, New England, Virginia, and the Carolinas. These thresholds highlight instances in which above-freezing temperatures exist in the low-level atmosphere near the surface, but below-freezing temperatures characterize the layers immediately above the surface.
A comparison of the precipitation-type estimates given by each algorithm with the corresponding observations is given in Fig. 1. This plot is representative of the other cases evaluated. Using the Ramer algorithm, freezing rain is indicated in only two cases, both of which are classified as a freezing mix (Fig. 1a), given the very narrow range of conditions (I < 0.04) that indicate pure freezing rain. With the modification, 7 of the 11 freezing rain occurrences are correctly classified. The improvement results from better differentiation between snow and freezing rain. The remaining four occurrences are classified as rain, given Ts > 0°C. The modified algorithm misclassified one snow observation as freezing rain. As was the case with the original Ramer algorithm, seven rain observations were classified as freezing rain. In all cases, the value of Ts given by the RUC initialization was <0°C.
c. Observation-adjusted precipitation-type algorithm
Although the fixed thresholds for IFL and Ts used in the modified Ramer algorithm represent the majority of freezing rain observations, in reality these thresholds vary with location and storm system. This variation is an artifact of the empirical nature of the thresholds as well as biases in the model initialization data. Assuming the thresholds within a specific freezing rain event are less variable, observations from hours just prior to the forecast period have the potential to be used to tune the thresholds on a case-by-case basis. For instance in Fig. 1b, increasing the Ts threshold to 0.2°C would result in correctly identifying two more freezing rain occurrences, without incorrectly including any more liquid precipitation events in the freezing rain category. Although physically unrealistic, this adjustment likely compensates for a warm bias in the RUC initialization for this hour.
Tuning of the IFL and Ts thresholds first required the interpolation of the hourly gridded model initialization data to Automated Surface Observing Systems (ASOS) locations. Multiquadric interpolation (Nuss and Titley 1994) was used, with the radius of influence set to 0.5° and 0.0025 specified as the smoothing parameter. The model Ts values at each grid within the radius of influence were first adjusted to the elevation of the station, based on the temperature lapse rates given by the model. Horizontal interpolation of Ts was then applied on this constant-elevation surface.
Although topographic differences within a grid may affect the height of the IFL above a particular point, this variable was not adjusted for elevation. Such an adjustment was not possible, since the model only specified the average vertical temperature profile across the grid. This average lapse rate could be used to adjust surface temperature, but the influence of topographic variation on the vertical temperature gradient itself was not known.
Station observations of valid precipitation type (i.e., reports of freezing rain at temperatures above freezing were ignored) were collected for the 6 h prior to the target forecast initialization time. For instance, to estimate precipitation type at 1200 UTC based on a 6-h lead-time forecast, the observations from 0100 to 0600 UTC were pooled. The default IFL and Ts thresholds were used if fewer than 20 station hours reported freezing rain. Otherwise, the maximum and minimum values of IFL and Ts corresponding to freezing rain observations were identified and used as the new thresholds.
Separate gridded surface temperature adjustments were also computed. Following horizontal interpolation of the model surface temperatures to the station locations, station biases (interpolation − observation) were computed. The resulting field of biases was then interpolated back to the grid points (using a larger 2° radius of influence to account for the coarser spatial density of the stations). This provided a means of adjusting the model surface temperatures alone based on the observations. These adjusted values were retained for a later evaluation.
3. Ice accretion model
Once the occurrence of freezing rain was determined by one of the above methods, estimates of ice accretion thickness on surface objects were calculated at each model grid point using the simple ice accretion model (Jones 1998). The uniform radial ice thickness on a cylinder, accumulated over the duration of a storm, is calculated by
where Req is the uniform radial ice thickness (mm), N is the number of hours of freezing precipitation, ρi is the density of ice (=0.9 g cm−3), ρ0 is the density of water (=1.0 g cm−3), P is the precipitation rate (mm h−1), V is the wind speed (m s−1), and W is the liquid water content (Wj = 0.067P0.846j).
For each hour (and precipitation-type algorithm) a high and low Req value was computed. The high ice accretion estimate considered both the occurrence of freezing rain and freezing mixed precipitation, while only those hours during which freezing rain was indicated were used for the low estimate. Hourly values were summed to give a total storm ice accumulation.
Ice accretion estimates based on the observation-adjusted precipitation-type algorithm were also adjusted based on conditions observed during the 6 h prior to forecast initialization. The gridded wind speed and precipitation values were interpolated to the ASOS locations using the multiquadric procedure. For wind speed, the logarithmic wind profile was used to first adjust the model winds to the anemometer height. These interpolated values were compared to the observed data creating a field of bias values at each station. Biases for wind speed and precipitation amount were defined as ratios (interpolated/observed). In the case of wind speed, the field of station biases was then interpolated back to the grid locations using the multiquadric procedure with a larger 2° radius of influence. This allowed the original gridded model data to be adjusted based on the recent bias experienced by nearby observations. For precipitation, a single domain-wide adjustment equal to the median of the individual station biases was retained. Gridded precipitation adjustments, analogous to those for wind speed, but without the elevation adjustment, were also computed for later evaluation.
4. Data and methods
a. WRF ice accretion forecasts
Hourly forecasts from the Weather Research and Forecasting (WRF) model (Skamarock et al. 2005) provided the input for each of the three precipitation-type algorithms and the simple ice accretion model. Two model domains, one centered on the border of North Carolina and Virginia (VANC) and the second over central New York State and northern Pennsylvania (NY), were considered. These areas corresponded to the service areas of collaborating electric utilities and are shown in the figures in subsequent sections. Within each domain, 12-km horizontal resolution WRF model forecasts with 31 vertical levels were made. These forecasts incorporated the following WRF model options: WRF single-moment (WSM) 3-class microphysics, Radiative Rapd Transfer Model (RRTM) longwave radiation, Dudhia shortwave radiation, Mellor–Yamada–Janjić (Eta) turbulent kinetic energy (TKE) scheme, and the RUC land surface model. The model boundaries extended 5° (latitude or longitude) beyond the forecast domains in all directions.
Prior to April 2002, 40-km RUC analyses (Benjamin et al. 1998) were used to initialize the WRF and specify boundary conditions. For subsequent dates, 20-km RUC analyses were available and used. In addition to the increased horizontal resolution, the 20-km analyses use hybrid isentropic-sigma vertical coordinates, while the coarser-resolution analyses use 50-mb isobaric vertical coordinates.
The WRF model was initialized every 6 h and produced hourly forecasts for the subsequent 12-h period. For each ice storm event the first initialization time was selected such that it preceded the first freezing rain report by at least 3 h. For each model run, 12 different radial ice thickness forecasts were produced. These reflected the three precipitation-type algorithms and the high (mixed freezing precipitation) and low (only freezing rain) options for computing ice accretion. In addition, these six values were computed using hours 1–6 of each forecast run and separately hours 7–12. These are referred to as 0- and 6-h lead forecasts, respectively.
b. Icing events
Icing events were identified using the queriable database of National Weather Service (NWS) storm reports available online (http://www4.ncdc.noaa.gov/cgi-win/wwcgi.dll?wwEvent/Storms). Hereafter, these are referred to as storm reports.
From this archive, all events listed as ice storms or freezing rain were identified and the textual accounts given by the reports were examined to quantitatively describe the total ice accumulation for the event. Overall, 23 ice storm events were identified that affected the VANC domain. An additional 15 cases were identified in NY. In both cases these storms occurred from January 1999 through February 2004. Of these events, 3 of the New York cases triggered a storm response by the New York State Electric and Gas Company and 10 of the events affected the collaborating utilities in Virginia and North Carolina. A list of the analyzed storms is given in Table 1.
Multiple storm reports often described the same large-scale storm event. This occurred primarily when a storm affected the forecast area of more than one NWS office. These separate reports were treated as subcases to account for spatial differences in ice accumulation (Table 1). Likewise, subcases also described instances where a single NWS forecast office reported different ice accumulations for groups of counties within its forecast area. Thus, some subcases corresponded to areas in which rain or snow, but no freezing rain, was reported. Subcases were classified as significant if more than 12.7 mm of ice accumulated or if the storm report described the storm as significant or major. This threshold is higher than the 6.0-mm value that typically triggers the issuance of winter storm warnings in the regions; however, it was necessary to assure an adequate sample of events with minimal ice accumulations.
c. Quantification of ice accretion observations
From the storm report ice accumulations, a single numeric ice accretion value was assigned to each subcase. The reports often specified a range of ice accumulation amounts over an area. In these cases the maximum value was retained. This provided consistency with other reports that described accumulations using terms such as “up to 0.5 in.”
Several problems are inherent to the ice accretion values obtained from the storm reports. Given the textual nature of the accounts, the reported accretions lacked precision. The majority of the reports are based on citizen reports to NWS offices and thus the reported values are not direct measurements, but most often visual estimates. The storm reports are also biased by factors such a population density and thus ice accretion may be unreported (or underestimated) in relatively unpopulated areas. Nonetheless, the storm reports are the most comprehensive source of information describing both the magnitude of icing and the spatial extent of storm impacts.
A more systematic and serious problem with the Storm Data ice reports is that they are unlikely to represent uniform radial ice thickness. Rather, the reports reflect either ice thickness on a horizontal surface or maximum ice radius or diameter on a wire or branch. In almost all cases, these measurements will exceed the value of Req given by Eq. (3). Thus, when directly comparing forecasted ice accretion radial thicknesses to the Storm Data reports of ice magnitude, this artifact will lead to a systematic apparent underprediction. To compensate for this imprecision, radial ice thickness values were also computed at ASOS stations within each WRF domain using observed data as input to the ice accretion model. These estimates provided a means of directly comparing the influence of the three precipitation-type algorithms, without the bias inherent in the Storm Data reports. However, given the relatively low density of ASOS sites, significant icing may go undetected if only this network is used for validation. These icing events are captured in the storm reports, providing qualitative validation data on both the spatial extent of the icing and the relative spatial variation in the ice thickness as characterized by the reports (as opposed to the true radial thicknesses).
Ice thickness estimates using ASOS observations and the ice accretion model are typically higher than those derived from ASOS icing sensors (Jones et al. 2004). These sensors have become operational at most ASOS sites and detect icing by sensing the ice mass on a cylindrical probe. Several factors contribute to the difference in ice accretion as reported by the sensors. The ice accretion model uses wind data from the 10-m level. The increase in wind speed with height gives a higher Req than would be obtained using winds at the lower ice sensor level. Errors in ASOS wind and precipitation observations (which may be more likely during freezing rain) can also bias the model estimates. The ice accretion from the model is further inflated when the mixed freezing precipitation is treated as freezing rain.
Supplemental ice accretion data were also available for two VANC storms. For the January 2004 storm a map of power outages, provided by Progress Energy, provided indirect evidence as to the spatial extent of the storm. Likewise, an outage plot provided by Dominion illustrated the spatial extent of the February 2003 ice storm in Virginia. The power outage maps highlight locations that were most affected by the storms, presumably due to relative maximum ice accumulation. Although nonmeteorological factors such as the effectiveness of tree-trimming programs and the age and maintenance history of the distribution system affect individual outages, the concentration of outages in areas of maximum ice accretion highlights the overwhelming influence of the meteorological factors. Liu et al. (2008) confirm this observation quantitatively. Jones et al. (2004) provide maps of the dominant observed precipitation type and ice accumulation (based on the ice accretion model) for the 4–5 December 2002 VANC storm that can be compared to similar maps generated with the WRF forecasts.
a. Comparison of lead times
Figure 2 compares total storm ice accumulations based on 0- and 6-h lead times with the storm report amount. A total of 89 subcases from NY and VANC were compared. In both cases, the modified Ramer precipitation-type algorithm is used and mixed freezing precipitation is assumed to be freezing rain. These results are similar to those for the original Ramer algorithm and low freezing rain estimates (i.e., ignoring freezing mix cases).
Collectively, there is little difference between the ice accretion forecasts based on the two lead times. The 6-h lead-time forecasts are less biased, but are slightly more variable. At zero lead time there is a tendency to underpredict ice accumulations. At a 6-h lead, eight forecasts overpredicted the observed ice accumulations by more than 15 mm. Only three of the 0-h lead forecasts had errors in this range. In four of the 6-h lead cases, this occurred with storms from the early part of the study period when only the coarser 40-km RUC initializations were present. Based on these comparisons, and the practical advantages of the longer lead time to the intended application, 6-h lead-time forecasts are used in subsequent analyses.
The remaining large differences at 6-h lead occurred during the 4–5 December 2002 VANC storm, primarily in the south-central North Carolina counties along the South Carolina border. In this area, Jones et al. (2004) show radial ice thicknesses that are in agreement with (but lower than) the 12.7-mm storm report maxima. However, the maximum 6-h lead forecast accumulations in this area generally exceed 30 mm. Based on the 0-h lead forecasts, accumulations of less than 20 mm were predicted in all but one case. In this region, Jones et al. (2004) show total liquid precipitation amounts in the range of 40–60 mm. The discrepancy appears to be related to the amount of liquid precipitation predicted by the WRF. Based on the precipitation received in four sequential 0-h lead forecasts (i.e., four nonoverlapping forecasts of rainfall in hours 1–6), the area receives up to 73 mm of precipitation. Using 6-h lead forecasts of rainfall (i.e., forecasts of rainfall in hours 6–12), nearly 100 mm of precipitation occurs in parts of the region. At a minimum (i.e., when the wind speed equals 0 m s−1), the difference between the 100-mm precipitation forecast and the 40–60 mm of precipitation reported by Jones et al. (2004) explains the difference in forecasted ice accretion.
b. Comparison of precipitation-type algorithms
Using the 6-h lead forecasts, Figs. 3 and 4 compare the forecasted radial ice thicknesses based on the original and modified Ramer precipitation-type algorithms with the storm reports for the NY and VANC subcases, respectively. Comparisons are also made using forecast ice accumulations based on only freezing rain (low) and mixed freezing precipitation (high). The comparisons are consistent and show that forecasts based on the original Ramer precipitation-type algorithm (RAM) consistently underpredict the observed ice accumulation. The modified Ramer precipitation-type algorithm (mRAM) also underpredicts the storm report amounts in most cases, but to a lesser degree. This consistent underprediction is partially an artifact of the bias in the Storm Data reports. For the NY cases, there is also a tendency for the mRAM predictions to be more variable; this feature is not apparent in the VANC cases.
As expected, the ice accretion estimates that include mixed freezing precipitation are higher than those based only on freezing rain. In some cases the differences are subtle, with only small changes in the median differences (Figs. 3 and 4). Increases in the median differences are largest for the RAM estimates in which several rain events are characterized as freezing mixed precipitation.
Underprediction is to be expected in these comparisons given the relationship between Req and the thicknesses reported in the storm reports. To remove the influence of the imprecision of the storm reports and their inherent biases, radial ice thickness estimates using observed and forecast data as input to the ice accretion model were compared at ASOS stations. In this comparison, some spatial uncertainty was introduced as the values using the ASOS observations represent a specific point, while the forecast values are an average for a 12-km grid encompassing the ASOS site. As the results based on the NY cases were analogous to those for VANC, only the VANC results are shown in Fig. 5. In these ASOS comparisons, the results are not differentiated by ice accretion amount, given the low number of ASOS ice estimates >12.7 mm. The results are again segregated into pre- and post-April 2002 periods, owing to the differences in the resolution of the available RUC initialization data.
When both the forecast data and observations indicate icing, the RAM and mRAM precipitation-type algorithms give ice accretion estimates that are similar to those based on the ASOS data (Fig. 5b). These general results are also repeated for cases in which the forecast data indicated icing, but none was reported at the ASOS site (Figs. 5c and 5d). In cases where the observations indicated icing but none was evident using the forecast data, the distributions of average errors were similar for all methods. Nonetheless, the smallest differences were associated with the mRAM algorithm (Figs. 5e and 5f). Likewise, the mRAM algorithm was associated with the fewest cases in which icing was detected by the ASOS, but not the model.
Overall, the ASOS comparisons corroborate those based on the storm reports. Icing forecasts using forecast data and the RAM precipitation-type algorithm tend to underestimate the observations. Although ice accretion forecasts based on the mRAM routine tended to be the less biased, there are cases where the difference between the forecast accumulations and the storm reports exceed 20 mm. Such case-specific examples are examined more closely in the next section to identify the causes for these differences other than the deficiencies of the Storm Data reports.
c. Spatial ice extent comparisons
Six case studies were used to illustrate the ability of the forecasting procedure to capture the spatial pattern and extent of ice accretion. The ice accretion events depicted in Figs. 6 and 7 were chosen since they were of interest in terms of the extent of the power outages associated with them. The areal coverage of these two storms is representative of the majority of the storms studied (Table 1). In Fig. 6, freezing rain fell in two separate periods from 25 to 27 January 2004. Storm reports indicated as much a 19 mm of ice accretion in northeastern South Carolina and southeastern North Carolina (Fig. 6a). Power outages were common in eastern North and South Carolina, with the majority of the outages clustered in the counties with the highest observed ice accretion. Mainly, rain fell along the immediate coast, while snow was the predominant precipitation type over western North Carolina. As much 12 mm of ice was reported in counties in extreme southwestern Virginia. Freezing rain impacts, automobile accidents, and scattered power outages were reported in the Washington, D.C., area, without mention of accumulation
In support of the boxplot comparisons (Figs. 3 –5), use of the RAM precipitation-type algorithm resulted in considerable underestimation of the extent and magnitude of the observed icing (Fig. 6b). Only 6 mm of radial ice accretion was forecast in the areas that received 19 mm of ice thickness. Although this difference may arise due to the nature of the Storm Data report data, the areas of maximum ice accumulation were coincident with the locations of the most severe impacts. Likewise, the RAM algorithm indicated freezing rain in both southwest Virginia and the Washington, D.C., area.
The freezing rain accretion forecast based on the mRAM precipitation was more in line with the observed impacts both in terms of the extent and magnitude of the icing (Fig. 6c). The maximum ice accretion was double that given by the RAM algorithm. However, the 12-mm maximum still underestimated the 19-mm observations, and the location of the maximum is shifted. The majority of counties reporting freezing rain impacts received at least 6 mm of ice accretion based on the mRAM procedure. In southwestern Virginia, the mRAM icing forecast was also more in line with the observations compared to that using RAM. The freezing rain forecasts for the Washington, D.C., area were similar.
Adjustment of the mRAM and ice accretion model thresholds and inputs (mRAMa) increased the freezing rain accretion forecasts to values that match those reported in the storm reports (Fig. 6d). The adjustment primarily affects the southeastern portion of the study area, increasing freezing rain accretion by 7 mm on average. Except for an expansion of the area of >12 mm ice accretion, the adjustment has little effect on the icing magnitude across central North Carolina or Virginia. Although the mRAMa forecast does capture the sharp ice accumulation gradient that was observed along the coast, the area of >19 mm ice accretion is displaced farther southeastward than is indicated by the impact reports.
Figure 7 shows a similar progression of increasing ice accretion totals among the precipitation-type algorithms. In this event, freezing rain fell during most of 27 February 2003 across central North Carolina and parts of central and eastern Virginia. With the exception of counties in extreme south-central area of the state (highlighted counties in Fig. 7a), snow was the main precipitation type across Virginia. As much as 15 cm of snow fell in central Virginia. To the east, 10 cm of snow was reported along with 3 to as much as 12 mm of ice accumulation. Power outage data for this case was limited to Virginia. As in Fig. 6, outages tended to cluster within the south-central Virginia counties with the highest observed ice accretion.
Unlike the case from Fig. 6, all of the forecast methods misplace the location of the maximum freezing rain occurrence (Fig. 7). The simple translation of the forecasted freezing rain areas by 150 km to the northeast would juxtapose the forecasted maximum over the areas with greatest impacts. In terms of maximum ice accretion amounts, the typical underestimation of ice accretion using the RAM algorithm is again illustrated as maximum ice accumulates are restricted to near 6 mm (Fig. 7b). The mRAM algorithm, however, offers little improvement with a similar ice accretion maximum and only a slight expansion of the area of 2–6-mm ice thickness (Fig. 7c). Only after adjustment using previous observations does the maximum forecasted ice thickness increase to 12 mm. While this value, and the area of >6 mm ice thickness, is in agreement with the majority of the storm reports, the highest (near 25 mm) ice thickness reports are still not captured by the forecast.
The misplacement of the axis of highest freezing rain totals in this event appears to result from the placement of the WRF precipitation field. In the model, the axis of heaviest storm total precipitation is aligned east–west along the southern North Carolina border (Fig. 8). This is in contrast to observations based on nearly 1000 rain gauges that show a precipitation axis aligned southwest to northeast coincident with the area of highest reported ice accretion (Fig. 8a). Adjusting the precipitation amounts based on the median precipitation bias across the forecast domain reduces the precipitation totals given by the model. The position of the heavy precipitation axis is not altered (Fig. 8c). Adjusting the precipitation totals on a grid-by-grid basis via spatial interpolation of the hourly rainfall observations (rather than using a single domain-wide adjustment) did little to influence the position of the freezing rain axis. This latter adjustment technique still does not capture the pattern of heaviest ice accretion for two reasons. First, to provide a true forecast, precipitation amount biases are based on observations from a period 6–12 h prior to the forecast. Therefore, even if the heavier precipitation is identified and adjusted, it will be assumed to occur at a point 6–12 h later in the storm, when other parameters may not be conducive to freezing rain occurrence. Also, given the spatial density of the available hourly rain gauge network and the relatively narrow daily heavy precipitation axis, the adjustment for the event is modest as gauges with little bias (from areas outside the precipitation maximum) drive the spatial interpolation. Altering the radius used to define the area within which stations influence the spatial interpolation of precipitation amount had little effect.
Several other methods of adjusting the forecast fields based on previous observations were also evaluated. These failed to improve to the ice accretion forecasts presented in Fig. 7d. Similarly, the accuracy of the ice accretion forecasts shown in Fig. 6, as well as those presented in subsequent figures, was either unchanged or diminished when these methods were applied. In addition to the grid-by-grid precipitation adjustment, these modifications included computing a domain-wide precipitation bias limited to those stations reporting freezing rain, and retaining the default ice fraction IFL thresholds, but basing the Ts threshold on the observation-adjusted temperature field. The IFL and Ts thresholds were also modified to define the smallest region in the IFL and Ts scatterplot that encompassed at least 80% of the total freezing rain observations and maximized the expression
where Z is the total number of freezing rain observations within the subregion and N is the number of nonfreezing rain precipitation types included in the subregion.
In Fig. 9, ice accretion forecasts for the December 2002 storm documented by Jones et al. (2004) are given. Figure 9a shows the extreme magnitude and extent of this icing event. Unlike the previous cases, the mRAMa algorithm produces the lowest ice accretion values (Fig. 9d) with maximum ice thicknesses of about half of the reported maxima. Nonetheless, the extent and general pattern of the icing are represented by the forecasts. The differences between the forecasts given by the mRAM and RAM algorithms are subtler. The RAM algorithm underestimates the observed ice thicknesses in southeastern Virginia, indicating amounts of <2 mm where 6–12 mm was observed (Fig. 9b). This feature is correctly forecasted using the mRAM algorithm (Fig. 9c). The two forecasts also differ in northwestern South Carolina and northeastern Georgia. In this region, where the observed ice thicknesses approached 38 mm, the mRAM algorithm gave forecast ice accretion in the range of 25–32 mm (Fig. 9c). Ice thickness forecasts based on the RAM algorithm did not exceed 25 mm (Fig. 9b). Both the RAM and mRAM overestimated ice accumulation in Georgia. This is likely an artifact of this area being near the boundary of the WRF domain.
In contrast to this widespread freezing rain event, Fig. 10 shows a localized event that occurred on 7 April 2003. The occurrence of freezing rain in this event was elevation dependent. Given the consistent underestimation of ice accretion by the RAM algorithm, results are not shown for the method nor are they presented for subsequent events. The mRAM and mRAMa algorithms give similar ice accretion forecasts, both of which are generally consistent with the storm reports. The mRAM forecasts tend to be higher than the reported ice accretion. Neither method captures the isolated reports of up to 12 mm of icing in several counties in western Virginia. Icing in these locations was driven primarily by elevation. Thus, the 12-km WRF resolution was not able to capture the finer-scale elevation features that experienced icing. Likewise, the mRAMa method depends on the ability of the available observations to replicate local topography. Given the sparse density of the station network, such features are not readily captured. Work to downscale the WRF ice accretion estimates to a finer spatial resolution is on going.
The forecast methods displayed similar characteristics when applied to cases that affected New York State. Figure 11 shows a mixed precipitation event that affected southern New York and northern Pennsylvania on 1–2 January 2003. The storm was mostly a freezing rain event; however, through the event precipitation transitioned from rain to freezing rain to snow at most locations. The change to snow was most pronounced in eastern New York. As in the VANC cases, both the mRAM and adjusted mRAM forecasts capture the spatial patterns of ice thickness indicated by the observations, with maxima in north-central Pennsylvania and southeastern New York. Without adjustment, the mRAM algorithm underestimates ice accretion across most of the domain, although to a much lesser degree than is given by the RAM algorithm (not shown). The adjusted ice thickness forecasts agree with the observations in most New York counties. However, the scattered 25-mm ice accretion reports in Pennsylvania are not captured by the forecasts. Presumably, this results from the smoothing both of the topographic features and the 12-km horizontal resolution model data to 0.2° for mapping. Several 12-km grids within the >12 mm contour in central Pennsylvania had forecast ice accretions totals between 15 and 20 mm.
The results of a second New York storm event occurring on 5 March 2003 are not shown graphically. In this event the occurrence of ice accretion was minimal as rain transitioned to light snow across most of eastern and central New York. Although freezing rain was reported in some locations, there were no storm reports of ice accumulation (other than slippery roads). For this event, ice accumulation forecasts using both the mRAM and mRAMa algorithms were less than 0.5 mm.
Forecasts of freezing rain accretion on wires and trees are an important component in the short-term decision making process of electric utilities faced with positioning human resources, prioritizing repairs, and minimizing the duration of power disruptions. Such forecasts are feasible, at the appropriate spatial resolution, using WRF model forecasts of precipitation amount, wind speed, and the thermodynamic profile of the lower atmosphere in combination with the ice accretion model described by Jones (1998) and a modified version of the Ramer (1993) precipitation-type algorithm.
Moreover, this combination of models can also be used to develop relatively high-resolution ice accretion climatologies, based on hindcasts generated from archived model initializations. Such information, when coupled with distribution system models, is critical to a range of electric utility decisions ranging from long-term planning to the economic evaluation of tree-trimming and maintenance programs. Current observation-based climatologies are not available at the needed spatial resolution. In addition, climatologies based on this approach have the potential to serve as a foundation for research focused on interannual variations in ice accretion as well as long-term trends in the severity and frequency of localized icing events.
Used in its original form, the Ramer algorithm consistently underestimated the magnitude of freezing rain accretion in nearly 40 freezing rain events of varying magnitudes that affected Virginia, North Carolina, New York, and Pennsylvania compared to the refined model. Modification of the model assessed the potential for freezing within the atmospheric layer defined by the surface and the lowest 0°C isotherm. This modification consistently increased the prevalence of freezing rain, resulting in ice accretion estimates that were more representative of both human estimates of observed ice thickness and values generated using ASOS observations as input to the ice accretion model. Adjustment of the forecasted variables based on the biases inherent in the forecast made 12 h previously, in general, improved the ice accretion forecasts. However, in some cases, such adjustments resulted in forecasts that were inferior to those without adjustment.
In its current form, the modified Ramer approach (either with or without adjustment) shows promise as a means of providing short-term ice accretion forecasts to the electric utility industry. Such forecasts must characterize both the magnitude and location of icing to be of practical use, which the 6–12-h forecasts evaluated here do with sufficient accuracy. It should be noted that, in all cases, the forecasts were able to segregate locations with accreting freezing rain from adjacent areas where snow and/or rain was the predominant precipitation type. This was true for events with widespread freezing rain accumulation as well as storms with limited (both spatially and temporally) freezing rain occurrence.
Two additional modifications to the forecast methodology are likely to lead to improvements in forecast accuracy and practical usefulness. Sophisticated models of the electric distribution infrastructure require icing data at approximately 1 km × 1 km resolution to adequately anticipate points of significant failures and estimate the length and magnitude of a power disruption. Based on the storm reports from the cases studied, it appears that elevation will be the likely factor influencing ice accretion at this scale. In lieu of directly obtaining WRF output at this resolution, which would require limiting the forecast domain to an area smaller than that serviced by a single utility, it is feasible that model lapse rate and surface temperature data could be used to downscale the 12-km data to the desired resolution. Assuming the horizontal gradients are minimal within the freezing rain shield, the lapse rate and elevation-adjusted temperature can be statistically interpolated to a finer digital elevation grid. The modified Ramer precipitation-type algorithm can then be applied to these interpolated variables to judge freezing rain occurrence. Ongoing research is exploring this methodology more formally.
Human forecaster intervention will provide the second means by which forecast accuracy can be improved. Such intervention would provide a means of accounting for forecast biases, such as the location of the heaviest precipitation, that are not addressed by the simple automated adjustments that were presented here. It should be noted that several other adjustment techniques, in addition to those presented in Fig. 8c, were evaluated in this work; however, none improved upon the ice accretion forecasts presented. These adjustments included 1) use of the default ice fraction thresholds, with observation-adjusted surface temperature data to specify precipitation type; 2) use of the smallest ice fraction and surface temperature ranges that encompassed at least 80% of the freezing rain observations; and 3) the computation of precipitation amount biases based only on those stations reporting freezing rain.
The methodology applied here could also be applied to longer-lead-time WRF forecasts. Although the 6–12-h forecasts evaluated here met the minimum forecast lead-time requirements of the utilities that were partners in this work, longer-lead forecasts of comparable accuracy could also be used. Such forecasts would need to rely solely on the model-derived ice accretion estimates, since adjustment based on observations more than 12 h old will have an adverse affect on many forecasts, particularly those associated with short-lived and/or localized ice accretion. The small median biases for our combined cases show promise for using these methods to create model-based ice storm climatologies to aid utilities in long-range planning.
This work was supported by NSF Grant CMS-0408525. Partial support was also provided by NOAA Contract EA133E-02-CN-0033. We are grateful to Yolanda Roberts for conducting many of the WRF ice accretion forecast runs.
Corresponding author address: Art DeGaetano, Northeast Regional Climate Center, Cornell University, 1119 Bradfield Hall, Ithaca, NY 14853. Email: email@example.com