William Burrows works in the Observation Based Research Section, Meteorological Research Division, of Environment and Climate Change Canada (ECCC). He has enjoyed a lengthy career with ECCC, first as an operational forecaster, then Instructor of Meteorology, and finally as a Research Scientist. For many years, he was located in Toronto. Since 2004, he has worked in Edmonton. His main area of expertise is designing and implementing statistical and machine-learning models that use post-processed output from Numerical Weather Prediction (NWP) models to forecast high impact weather elements. His models include stratospheric ozone, ground-level ozone, lightning, lake-effect snow, fog/stratus, and blizzards.
Curtis Mooney works for ECCC in the Meteorological Service of Canada’s National Lab-West. He joined ECCC nearly two decades ago, after spending time in the private sector doing air quality assessments for industrial clients. With ECCC, he first worked as an Arctic region operational forecaster, then as an instructor with the Meteorologist Operational Internship Program (MOIP). His expertise is in meteorological and air quality modelling. He is involved in projects related to the development and transfer to operations of forecast techniques for fog and stratus, blizzards, and precipitation typing.
In this contribution to Polar Prediction Matters, William Burrows and Curtis Mooney discuss blizzard conditions in recent years in the Canadian Arctic and some automated blizzard prediction products they have developed. The authors are co-located with the Prairie and Arctic Storm Prediction Centre (PASPC) and the Canadian Meteorological Aviation Centre – West (CMAC-W) in Edmonton, Alberta, Canada. Their prediction products have been developed over a period of years in consultation with forecasters in PASPC and CMAC-W, and are often used in operations.
Blizzards are a regular occurrence from October through May in the Canadian Arctic and constitute one of the major forecasting challenges for meteorologists tasked with predicting Arctic weather. The combination of low temperature, strong wind, and reduced visibility to near 0 km in blowing snow makes a blizzard one of the most disruptive and hazardous of Arctic weather events. Blizzard conditions last anywhere from a few hours to more than a week in Canada’s Arctic, and are particularly common north of the treeline (northernmost extent of the boreal forest) due to the open terrain there. The Meteorological Service of Canada (MSC) considers blizzard conditions to be temperature below 0°C, wind 40 km/h or stronger, and visibility ≤ 1/4 statute miles (about 0.4 km) in blowing snow or concurrent precipitating snow and blowing snow. MSC will issue a blizzard warning 18 hours in advance if blizzard conditions are expected to last 4 hours or more south of the treeline and 6 hours or more north of the treeline. Figure 1 is a good example of what the local environment looks like during a blizzard.
Here, we are showing some of the automated research products we have developed that run in real time to aid forecasters in predicting where and when blizzard conditions are likely to occur.
The Blizzard Environment
The main environmental factors for blizzard conditions are below freezing temperature, high wind speed in the lowest few hundred meters near the ground, a source of snow either as precipitation or as snow available for lifting off the ground, and the condition of the snowpack (loose or hard-packed). Figure 2 shows a vertical sounding taken at 12 UTC on 29 March 2018 when a radiosonde balloon was released during blizzard conditions at Baker Lake, located west of Hudson Bay (for location see Fig. 3). The visibility at the time of this sounding was very likely near 0 km in precipitating snow and/or blowing snow. Only an automatic machine instrument (AUTO) was reporting at the time because the observer had ceased taking manned observations several hours earlier when conditions became extreme. For those with a meteorological background some signature features typical of blizzard conditions are seen in Fig. 2: 1) low temperature near the ground; 2) a nearly moist adiabatic lapse rate of wet-bulb potential temperature extending from the ground to the base of a strong temperature inversion located a small distance above the ground; 3) strong low-level wind with little direction change in a relatively deep layer. The nearly moist adiabatic lapse rate above the ground is caused by snow sublimating into the air after it is lifted off the ground by mechanical turbulence in the strong wind.
Frequency of Blizzard Conditions in the Canadian Arctic
To assess how often blizzard conditions occur across the Canadian Arctic, we have analyzed observations from a number of observing stations that regularly document weather observations in so-called METeorological Aerodrome Reports (METARs). Figure 3 shows the percentage of hours where blizzard conditions were reported as defined above in METARs in the October through May periods in the years 2014-2018. Results from these four years take into account that many hours of observations are routinely missing because most AUTO reports from Canadian Arctic stations do not include visibility; observations at most stations revert overnight to “AUTO” machine reports, and in bad conditions during the day, a manned observer may stop observing and turn on the machines. Figure 3 reveals that blizzard conditions are most frequent at sites east and north of the treeline, particularly on the western coast of Hudson Bay and inland from there. This area is mainly flat open terrain, and is well known to Canadian forecasters as “blizzard alley”. It is noteworthy that many Arctic observing sites lie in sheltered locations, particularly in the rugged terrain of the eastern Arctic, where blizzard conditions may not be seen at a site when they are occurring in the nearby open country. Dewar Lakes, located on high open terrain in the middle of Baffin Island, recorded a much higher frequency of blizzard conditions than other stations on the island.
Figure 4 shows the percentage of all hours with blizzard conditions that were classified as “clear-sky”, that is, where blowing snow only (no precipitating snow) was the reason for reduced visibility (see also Figure 1). The importance of predicting low-level wind accurately in the Arctic when wind speeds are sufficient to lift snow from the ground is evident. The percentage of clear-sky blizzard conditions reaches around 80% or more on the west side of Hudson Bay and at Alert on the northern end of Ellesmere Island.
Blizzard Prediction Products
Since blizzard conditions are determined from a combination of weather elements, output from a numerical weather prediction (NWP) model must be post-processed in order to derive products to forecast blizzard conditions. We have developed three products, each obtained by application of a different methodology. These products predict the environment where blizzard and near-blizzard conditions are likely, rather than predicting the exact wind speed and visibility. In order to cover short range to medium range prediction times, all three products run with output from current Canadian Meteorological Center (CMC) operational NWP models: global (GDPS), regional (RDPS), and the new high-resolution Arctic model (CAPS) that became available in January 2018. The model resolutions are: CAPS 3 km, RDPS 10 km, and GDPS 25 km. The GDPS and RDPS forecasts mainly cover the North American Arctic domain while the CAPS forecasts cover a circumpolar domain extending from the North American sector across the pole into northern Europe and Russia. We generate forecasts daily after the 00 UTC and 12 UTC model run times valid each hour for the 1-48 hour forecasts produced with RDPS and CAPS output, each hour for the 49-84 hour forecasts produced with RDPS output, and each 3 hours for the 87-120 hour forecasts produced with GDPS output. Forecasts for 3-5 day periods are especially important in the Arctic for planning surface travel as most occupied sites in the Arctic are isolated. We extend the forecasts over water bodies that are more than 80% ice covered since travel over ice covered water bodies is common in some Arctic regions. Forecasts are displayed in real-time on an internal Environment Canada research website http://weg-hal-fe01.edm.ab.ec.gc.ca/WxProducts/HAL_Winter accessible to all Canadian forecast offices. CAPS model forecasts are also uploaded to the CMC Collaboration website at http://collaboration.cmc.ec.gc.ca/science/rpn/YOPP and can be accessed by users outside of Environment Canada; the public release of these forecasts is one of ECCC’s contributions to the ongoing Year of Polar Prediction (YOPP), initiated by the World Meteorological Organization to enhance polar environmental predictions.
Based on the output of these forecast systems, we derive three products: The first product, known as the Blizzard Potential (BP), is derived from a set of forecaster expert’s rules to identify areas where blizzard conditions could develop over land and ice-covered areas with at least 1 cm of snow on the ground. The BP product is intended as “heads-up” guidance to alert forecasters a few days in advance to regions where blizzard conditions might occur. Wind, temperature, and precipitation are included in the rules but snow pack condition is not.
The second product, adapted from work by Baggaley and Hanesiak (2005), is a forecast of the probability of visibility ≤ 1 km due to blowing snow or to concurrent precipitating snow and blowing snow. Probabilities were obtained by analyzing nearly 40 years of surface observations of visibility, 10 m wind speed, and surface temperature from which simple equations were derived to predict the probability of blizzard conditions in both blowing snow only and in concurrent precipitating snow and blowing snow. The number of hours since the last snowfall is included as a proxy for snowpack condition. We apply the equations in real time to NWP model output to produce a probability forecast. This forecast production method, known as the “perfect prognosis” method, assumes the NWP model predictions are correct and does not account for NWP model prediction errors.
The third product we have developed is a prediction of the likelihood of blizzard conditions. For this we built a prediction model from a learning data set constructed by matching blizzard conditions in archived weather observations for two winter periods (October to May 2015-2017) with a set of predictors calculated from NWP model output that are physically related to the environment in blizzard conditions. A prediction model was generated from the learning data using Random Forests (Brieman, 2001), a powerful machine-learning algorithm for classifying events. This forecast production method, known as the “model output statistics” method, does account for NWP model prediction errors.
To illustrate these products, we show predictions for a recent Arctic blizzard event. Figure 5 is a 108-hour forecast valid at 12 UTC 29 March 2018 of the Blizzard Potential produced from GDPS model output. The overlaid winds are the average wind in the lowest 6 GDPS model levels from the ground upwards, which covers a layer approximately 500 meters deep above the ground. Five areas of BP categories are shown in order of increasing severity. Category 1 (green) is typically associated with possible drifting snow or blowing snow, category 2 (blue) with drifting snow and possible blowing snow, category 3 (orange) with near-blizzard to blizzard conditions in blowing snow, category 4 (yellow) with near-blizzard to blizzard conditions in precipitating snow and/or blowing snow, category 5 (red) with blizzard conditions in precipitating snow and/or blowing snow. The orange and red areas to the northwest of Hudson Bay in Fig. 5 are an early warning that severe winter conditions are predicted to develop there in a few days, and close attention to this region by forecasters is warranted.
Figure 6 shows an 84-hour forecast of the probability of visibility ≤ 1 km due to blowing snow or concurrent precipitating snow and blowing snow using the Baggaley and Hanesiak equations, valid at the same time as the Blizzard Potential forecast in Fig. 5. The forecast is produced with output from CMC’s RDPS model. The probability continues to be very high that widespread blizzard conditions will be occurring west and northwest of Hudson Bay in 3½ days according to this prediction.
Figure 7 shows a 48-hr forecast of the “yes vote ratio” for the likelihood of blizzard conditions valid at the same time as the forecasts in Figs. 5 and 6. This forecast is produced by our machine-learning model with output from CMC’s RDPS model. The meaning of the yes vote ratio can be explained as follows: our Random Forest model is comprised of an ensemble of 100 different decision trees constructed from the learning data, each predicting whether blizzard conditions will or will not occur at a grid point when presented with the forecast value of each predictor in the set of predictors used to construct the trees. The yes vote ratio is simply the fraction of the 100 trees that decided “yes, blizzard conditions will occur” at each grid point. The orange and red areas in Fig. 7 show where a majority of the trees voted “yes”. We see that blizzard conditions are still predicted over a large land region west and northwest of Hudson Bay as well as over several smaller areas.
As time draws closer to an event, a high-resolution model becomes useful for providing a detailed forecast. Even though our Random Forest model was built with learning-data predictors calculated from RDPS model output, we found it performs fairly well for predicting blizzard conditions when the predictor values are calculated from CAPS output data. Figure 8 shows a 24-hour forecast of the yes vote ratio using CAPS model output to drive our machine-learning blizzard prediction model. There is more detail in the forecast compared to a forecast from the RDPS model (Fig. 7). The area with the highest yes vote ratios for blizzard conditions is still located west and northwest of Hudson Bay, but the area is smaller. The synoptic weather system responsible for the blizzard conditions is better defined than in the 48-hour forecast shown in Fig. 7. Some of the differences may be due to the shorter prediction time of the forecast in Fig. 8 compared to Fig. 7, but the higher resolution of the CAPS model results in a more detailed prediction.
Figure 9 shows observations from the few stations located around the northern Hudson Bay area. Only those whose winds were ≥ 18 knots are shown since a lower wind speed does not trigger a warning for blizzard conditions. None of the stations reporting lower wind speeds were in the area of interest northwest of Hudson Bay. Stations reporting visibility ≤ ¼ statute mile are colored red. The observations are in good agreement with the forecasts discussed above.
The results shown in this article are a shortened version of a much more detailed paper we are currently writing for publication in a scientific journal.
We have shown here a brief climatology of blizzard conditions in the Canadian Arctic and a case example of the three automated products that we produce for forecasting blizzard conditions, which are driven by the CMC Global, Regional, and CAPS models. A new product which we are currently developing is a “Venn diagram” combination of all three products to highlight in a single figure the most likely areas that are predicted to see blizzard conditions.
We expect these blizzard forecast products would be useful to those who provide forecast services to a variety of users who live or operate in the north since they cover the entire Arctic region from near term to five days. People who live in specific locations are always interested in forecasts of high impact weather such as blizzards for their local area. Moreover, groups who need to know about potential harsh weather conditions over large portions of the Arctic will likewise benefit from these products. Such groups include agencies involved in aviation, search and rescue, military and civil defence, policing, and those who travel for hunting, visiting, or commerce between communities.
We welcome feedback as well as suggestions for improvements and collaborations.
Baggaley, D. G. and J. M. Hanesiak, 2005: An empirical blowing snow forecast technique for the
Canadian Arctic and Prairie Provinces. Wea. Forecasting, 20, 51-62.
Brieman, L., 2001: Random Forests, Machine Learning, 45, 5–32.