Real-Time Forecasting of Snowfall Using a Neural Network

Abstract

Abstract A set of 53 snowfall reports was collected in real time from the 2004/05 and 2005/06 cold seasons (November–March). Three snowfall-amount forecast methods were tested: neural network, surface-temperature-based 676-USDT table, and climatological snow ratio. Standard verification methods (mean, median, bias, and root-mean-square error) and a new method that places the forecasts in the context of municipal snow removal, and introduces the concept of forecast credibility, are used. Results suggest that the neural network method performs best for individual events, owing in part to the inverse relationship between melted liquid equivalent and snow ratio; hence, the ongoing difficulty of producing accurate forecasts of melted equivalent precipitation (a problem in all seasons) is compensated for rather than amplified when converting to snowfall amounts. This analysis should be extended to a larger selection of reports, which is anticipated in conjunction with efforts currently ongoing at the National Oceanic and Atmospheric Administration’s Hydrometeorological Prediction Center.