Consensus Probabilistic Forecasting

To make optimal decisions, end-users and decision support systems require information accurately describing the uncertainty of the underlying weather forecasts. Probabilistic forecast information about surface weather variables such as temperature, humidity, and wind speed will ideally take the form of a probability density function (pdf). Ensemble methods provide a natural approach for creating these types of forecasts. However, to create meaningful forecasts using ensemble methods generally requires production of a large number of realizations of a model forecast, which can be expensive in time and other resources. Moreover, calibration of the ensemble forecasts is often a concern.

As an alternative, a statistical approach is proposed in which forecasts generated by individual models are interpreted statistically to generate individual pdfs for the variable of interest (e.g., temperature). The resulting forecast distributions are then combined using weights based on forecast performance. This weighting procedure allows generation of multimodal forecast distributions. This approach is a probabilistic extension of the DICast system, a successful automated “traditional” scalar forecasting system.

Results are presented for temperature pdfs produced using output from two operational numerical weather prediction models (the Eta and GFS). The pdfs are evaluated using standard metrics such as the Continuous Ranked Probability Score and its components. This approach allows an evaluation of overall performance as well as performance for a variety of “events” ranging from extreme to average. Reliability diagrams and rank histograms provide additional insights into the forecast quality. Evaluation of the pdfs generated by these models indicates that they provide relatively reliable and skillful forecasts when compared to the deterministic forecasts and simple probabilistic forecasts based on climatology. In addition, the rank histograms indicate that the forecast spread is generally approximately correct. Integration of the pdfs provided by the two forecasting systems results in notable improvements, particularly with respect to the resolution and reliability of the forecasts. This improvement is associated with the ability of the integration process to apply larger weights to the forecast model that is providing the best performance.

Application of this method for temperature forecasts based on the output of only two NWP models produced encouraging results; use of additional models or model realizations would be expected to show additional capabilities. Evaluations of forecasts for other variables such as wind speed are in progress; initial results indicate that application of this approach is also beneficial for these forecasts.