Abstract
A suite of different artificial neural networks (ANN) is used to create 24-hour point forecasts for U.S. cities selected in the WxChallenge national forecast competition. Maximum and minimum temperature, maximum sustained wind speed, and total precipitation forecasts are taken across five weather forecasting models as input to the neural networks. The five models serving as input are: Global Forecast System (GFS) and North American Mesoscale (NAM) model output statistics (MOS), grid interpolated forecasts from the Hi-Res Rapid Refresh (HRRR) model, National Blend of Models (NBM), and National Weather Service point forecast matrix (PFM) forecasts. The ANNs are trained on over two years of historical model forecast data for every city. All ANNs—and weighted and unweighted ensembles of them—are compared against a multiple linear regression model and human forecasters to determine if ANNs provide an increase in skill of short-range forecasting of the given parameters. Performances are evaluated through 10 cities (80 forecasts) of the 2020–21 WxChallenge season.
How to Cite
Hill, A., (2021) “Comparing the Performances of Neural Network Architectures on Short- Range Weather Forecasts”, Capstone, The UNC Asheville Journal of Undergraduate Scholarship 34(1).
4
Views
2
Downloads