Uncertainty Quantification of Weather Forecasts

• Activity Number: 199 - SPEED: Data Expo
• Type: Contributed
• Date/Time: Monday, July 30, 2018 : 11:35 AM to 12:20 PM
• Sponsor: Section on Statistical Computing
• Author(s): Yu Wang* and Gong Zhang and Boyi Hu and Ho Yin Ho
• Companies: University of British Columbia and University of British Columbia and University of British Columbia and University of British Columbia
• Keywords: Kullback-Leibler divergence; Spatiotemporal modelling; Kriging; Uncertainty quantification; Spatial autocorreltion

Abstract

We consider the problem of evaluating how the performance of weather forecasts change as the forecasted date gets closer. We use a dataset containing weather forecasts and historical records of 113 cities in the USA over 3 years. The focus is on characterizing the accuracies of the forecasted min, max temperature, probability of precipitation given other associated environmental measurements. The forecasts data are based on an average over the whole area while the historical records are measured at the closest airport. To measure forecasting accuracies for min and max temperature, we use the differences between the forecasted data and the historical records, corrected by the distance between the forecasting area and the airport. To assess the accuracy of the predicted probability of precipitation (POP), we use the KL distance between the conditional distribution of observed precipitation given the predicted POP and the binomial distribution with the probability of success being the predicted POP. For uncertainty quantification, kriging and spatial autocorrelation models are applied on each proposed accuracy measures with historical environmental factors as additional predictors.