106 Assessment of Post-Processing Methods for Daily High and Low Temperature Prediction

Monday, 23 January 2017
4E (Washington State Convention Center )
Maxwell A Gawryla, The Weather Company, an IBM Business, Andover, MA; and J. K. Williams, P. P. Neilley, J. P. Koval, and S. Marshall

The Weather Company, an IBM Business, provides billions of automated forecasts to consumers every day, predicting many weather-related variables across the globe. Most of these variables, including temperature, are generated on a sub-hourly to hourly level through a system of statistical equations. In order to ensure consistency, daily high and overnight low temperatures are derived from the hourly temperature forecasts in a static post-forecasting stage. The purpose of this study was to compare several post-processing methods and their improvements in the overall skill of forecasting high and low temperatures without a substantial increase in necessary computational power or difficulty of implementation or maintenance.

The analysis used hourly forecasts from July 2015 through July 2016 from several hundred observing locations across the continental United States. Several methods of deriving a daily and overnight temperature were created and tested against METAR observations. Three methods used only the forecasted hourly temperatures: (1) a “bias” adjustment from the high (low), (2) a decision tree with adjustments that depended on time of day and forecast temperature range, and (3) extensions that included both linear and nonlinear functions. Additionally, the use of forecasted values of non-temperature variables were tested and evaluated. Skill was assessed through several metrics including bias calculation, root mean squared error, and percent within plus or minus 3°F. Preliminary results indicate that the more complex methods can provide noticeable improvement over simple “bias” adjustments to the predicted hourly time series.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner