Probabilistic Forecasting of Snowfall Amounts Using a Hybrid between a Parametric and an Analog Approach

Abstract

Abstract Forecast uncertainty associated with the prediction of snowfall amounts is a complex superposition of the uncertainty about precipitation amounts and the uncertainty about weather variables like temperature that influence the snow-forming process. In situations with heavy precipitation, parametric, regression-based postprocessing approaches often perform very well since they can extrapolate relations between forecast and observed precipitation amounts established with data from more common events. The complexity of the relation between temperature and snowfall amounts, on the other hand, makes nonparametric techniques like the analog method an attractive choice. In this article we show how these two different methodologies can be combined in a way that leverages the respective advantages. Predictive distributions of precipitation amounts are obtained using a heteroscedastic regression approach based on censored, shifted gamma distributions, and quantile forecasts derived from them are used together with ensemble forecasts of temperature to find analog dates where both quantities were similar. The observed snowfall amounts on these dates are then used to compose an ensemble that represents the uncertainty about future snowfall. We demonstrate this approach with reforecast data from the Global Ensemble Forecast System (GEFS) and snowfall analyses from the National Operational Hydrologic Remote Sensing Center (NOHRSC) over an area within the northeastern United States and an area within the U.S. mountain states.