Consensus Probabilistic Forecasting

To make optimal decisions, end-users of decision support systems require information accurately describing the uncertainty of the underlying weather forecasts. Air temperature, dew point temperature, and wind speed are critical surface weather variables in many economic sectors. The generation of sharp and calibrated probabilistic forecasts and their effective presentation to decision makers are current research challenges. This paper addresses the first of these challenges by describing an operational probabilistic forecast system developed at NCAR/RAL. This system is a probabilistic extension of DICast, an automated consensus forecast system which serves as the operational backbone for several major US weather providers. Today, forecasts of these continuously-valued weather surface variables are commonly generated and presented as deterministic (scalar) values. For example, the maximum temperature 2 days from now will be (exactly) 25oC. A more complex forecast representation is required describe the uncertainty in the forecast. Rather than forecasting a scalar, a probabilistic forecast ideally takes 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. The DICast probabilistic forecast system considers multiple numerical weather model inputs and uses a multi-model or “poor man’s” ensemble approach. Each model is interpreted statistically to generate individual pdfs for the variables of interest (e.g., temperature). The system then combines the resulting forecast distributions using weights based on the past forecast performance of each of the models’ pdf forecasts. The resultant consensus forecast is again a pdf. This weighting procedure allows generation of multimodal forecast distributions. The main conceptual difference between this probabilistic forecast system and the “traditional” scalar DICast system is that the probabilistic system produces and combines pdf’s rather than scalars. It is well understood that the combination of scalar forecasts produces statistically superior forecasts. The goal of this paper is to demonstrate that the same can be true for probabilistic forecasts.