Probabilistic temperature forecasting: a proof of concept

Communicating uncertainty in weather forecasts is becoming more important as decision makers become better able to apply uncertainty in their businesses practices and operations. Probabilistic precipitation forecasting has effectively communicated useful uncertainty for many decades. The National Weather Service Forecast Office in Tulsa, Oklahoma (WFO Tulsa) has been studying how probabilities may be applied to temperature forecasting and how those uncertainties may be communicated to customers. A proof of concept has been developed.

Thirty years of maximum and minimum temperatures were used to compute means and standard deviations for each day of the year at Tulsa International Airport. These temperature data follow a Gaussian distribution with the expected seasonal differences in variance: larger in the winter and smaller in the summer. These climatological distributions are compared to daily probabilistic maximum and minimum temperature forecasts based on a variety of MOS (Model Output Statistics), model and local WFO Tulsa forecasts.

Probability density functions of maximum and minimum temperature were computed separately for the official forecasts issued by the WFO for Tulsa International Airport and also for a variety of MOS guidance values based on several numerical weather prediction models run by the National Centers for Environmental Prediction and direct model output from that run by the European Centre for Medium Range Forecasts. Approximately ten years of forecasts and observed temperatures were used. The individual density functions were combined to create a final probability distribution of forecast maximum and minimum temperature. When the day-to-day changes in the temperature forecast are small, the variance of the combined distribution is small. Similarly, when the day-to-day changes in the temperature forecast are large, the variance of the combined distribution is large. Also, when MOS and model forecast vary widely, the combined density functions will also have large variances and can be multi-modal.

Sample results will be shown to demonstrate how the individual forecast distributions may be combined to provide improvement over the climatological distributions, and convey uncertainty for a variety of events.