Detecting Forecast Error Signatures

Abstract:

Incorrect weather forecasts can be deadly and very costly, so in this work we explore where and why statistical weather forecasting makes errors. Our weather forecasts come from the Global Forecasting System (GFS) Model Output Statistics (MOS). MOS is a statistical process that is based on a linear regression of a numerical model forecast tuned against observational data. The Global Historical Climatology Network (GHCN) is another database comprised of ground observations from weather stations around the world. For our research we concentrated on stations in the United States. As a case study, we computed the differences between the MOS forecasts and GHCN observations for 2018 and 2019. A K-means clustering algorithm was used to understand if there were common patterns of errors in similar types of stations. By looking at this data in a time series we understood where there were reoccurring errors. With this data we can calibrate the model to make better predictions.