Validation of Cool-Season Precipitation Forecasts at a High-Elevation Site in Utah's Little Cottonwood Canyon

Contemporary snowfall forecasting typically requires three components: a quantitative precipitation forecast (QPF), precipitation type identification, and application of a snow-to-liquid ratio (SLR). The components can vary significantly in the complex terrain of the western continental United States (CONUS), necessitating fine-scale forecasts of each component. SLR is particularly difficult to forecast due to atmospheric and terrain influences, intrastorm variability, and model limitations (e.g., inadequate resolution). Using high quality liquid precipitation equivalent (LPE) and snowfall observations collected during the 2019/20 to 2022/23 cool seasons (November 1 – April 30) at Alta, Utah, this presentation examines the performance of global, regional, and downscaled operational numerical model precipitation forecasts, operational SLR forecast methods used by the National Weather Service (NWS) for the National Blend of Models (NBM), and machine learning (ML) SLR forecast methods developed by the University of Utah.

Alta is located at the end of Utah State Route 210 in upper Little Cottonwood Canyon, which is bisected by 50 avalanche paths and has the highest uncontrolled avalanche hazard index of any major road in the world. Based on observations collected at 12-h intervals, mean cool-season snowfall and LPE at the Alta-Collins observing site used in this study are 1145 cm and 953 mm, respectively. During the study period, 12-h QPFs produced by the Global Forecast System (GFS) and High-Resolution Rapid Refresh (HRRR) exhibit negative mean bias errors, indicating an overall underprediction of LPE by 33% and 28%, respectively, with variations from storm-to-storm and season-to-season. Despite differing resolutions and terrain representations, the GFS and HRRR produce comparable 12-h QPF critical success indices at Alta for 50^th, 75^th, and 90^th percentile observed LPE events. SLR mean absolute errors for snowfall events are generally lowest for the University of Utah ML SLR methods (~3.6 – 4) and highest for the NBM operational SLR methods used by the NWS (~4.4 – 6). University of Utah ML SLR methods were trained using observations prior to the validation period. These results indicate the potential for machine learning applied to high-quality snowfall datasets to improve SLR forecasting over current operational NWS SLR forecast methods.